dtsi: rk3368: add uart & jtag pcl func for sdmmc mux
[firefly-linux-kernel-4.4.55.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) Nadia Yvette Chambers, April 2004
4  */
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/page-isolation.h>
25
26 #include <asm/page.h>
27 #include <asm/pgtable.h>
28 #include <asm/tlb.h>
29
30 #include <linux/io.h>
31 #include <linux/hugetlb.h>
32 #include <linux/hugetlb_cgroup.h>
33 #include <linux/node.h>
34 #include "internal.h"
35
36 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
37 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
38 unsigned long hugepages_treat_as_movable;
39
40 int hugetlb_max_hstate __read_mostly;
41 unsigned int default_hstate_idx;
42 struct hstate hstates[HUGE_MAX_HSTATE];
43
44 __initdata LIST_HEAD(huge_boot_pages);
45
46 /* for command line parsing */
47 static struct hstate * __initdata parsed_hstate;
48 static unsigned long __initdata default_hstate_max_huge_pages;
49 static unsigned long __initdata default_hstate_size;
50
51 /*
52  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53  */
54 DEFINE_SPINLOCK(hugetlb_lock);
55
56 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
57 {
58         bool free = (spool->count == 0) && (spool->used_hpages == 0);
59
60         spin_unlock(&spool->lock);
61
62         /* If no pages are used, and no other handles to the subpool
63          * remain, free the subpool the subpool remain */
64         if (free)
65                 kfree(spool);
66 }
67
68 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
69 {
70         struct hugepage_subpool *spool;
71
72         spool = kmalloc(sizeof(*spool), GFP_KERNEL);
73         if (!spool)
74                 return NULL;
75
76         spin_lock_init(&spool->lock);
77         spool->count = 1;
78         spool->max_hpages = nr_blocks;
79         spool->used_hpages = 0;
80
81         return spool;
82 }
83
84 void hugepage_put_subpool(struct hugepage_subpool *spool)
85 {
86         spin_lock(&spool->lock);
87         BUG_ON(!spool->count);
88         spool->count--;
89         unlock_or_release_subpool(spool);
90 }
91
92 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
93                                       long delta)
94 {
95         int ret = 0;
96
97         if (!spool)
98                 return 0;
99
100         spin_lock(&spool->lock);
101         if ((spool->used_hpages + delta) <= spool->max_hpages) {
102                 spool->used_hpages += delta;
103         } else {
104                 ret = -ENOMEM;
105         }
106         spin_unlock(&spool->lock);
107
108         return ret;
109 }
110
111 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
112                                        long delta)
113 {
114         if (!spool)
115                 return;
116
117         spin_lock(&spool->lock);
118         spool->used_hpages -= delta;
119         /* If hugetlbfs_put_super couldn't free spool due to
120         * an outstanding quota reference, free it now. */
121         unlock_or_release_subpool(spool);
122 }
123
124 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
125 {
126         return HUGETLBFS_SB(inode->i_sb)->spool;
127 }
128
129 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
130 {
131         return subpool_inode(file_inode(vma->vm_file));
132 }
133
134 /*
135  * Region tracking -- allows tracking of reservations and instantiated pages
136  *                    across the pages in a mapping.
137  *
138  * The region data structures are protected by a combination of the mmap_sem
139  * and the hugetlb_instantion_mutex.  To access or modify a region the caller
140  * must either hold the mmap_sem for write, or the mmap_sem for read and
141  * the hugetlb_instantiation mutex:
142  *
143  *      down_write(&mm->mmap_sem);
144  * or
145  *      down_read(&mm->mmap_sem);
146  *      mutex_lock(&hugetlb_instantiation_mutex);
147  */
148 struct file_region {
149         struct list_head link;
150         long from;
151         long to;
152 };
153
154 static long region_add(struct list_head *head, long f, long t)
155 {
156         struct file_region *rg, *nrg, *trg;
157
158         /* Locate the region we are either in or before. */
159         list_for_each_entry(rg, head, link)
160                 if (f <= rg->to)
161                         break;
162
163         /* Round our left edge to the current segment if it encloses us. */
164         if (f > rg->from)
165                 f = rg->from;
166
167         /* Check for and consume any regions we now overlap with. */
168         nrg = rg;
169         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
170                 if (&rg->link == head)
171                         break;
172                 if (rg->from > t)
173                         break;
174
175                 /* If this area reaches higher then extend our area to
176                  * include it completely.  If this is not the first area
177                  * which we intend to reuse, free it. */
178                 if (rg->to > t)
179                         t = rg->to;
180                 if (rg != nrg) {
181                         list_del(&rg->link);
182                         kfree(rg);
183                 }
184         }
185         nrg->from = f;
186         nrg->to = t;
187         return 0;
188 }
189
190 static long region_chg(struct list_head *head, long f, long t)
191 {
192         struct file_region *rg, *nrg;
193         long chg = 0;
194
195         /* Locate the region we are before or in. */
196         list_for_each_entry(rg, head, link)
197                 if (f <= rg->to)
198                         break;
199
200         /* If we are below the current region then a new region is required.
201          * Subtle, allocate a new region at the position but make it zero
202          * size such that we can guarantee to record the reservation. */
203         if (&rg->link == head || t < rg->from) {
204                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
205                 if (!nrg)
206                         return -ENOMEM;
207                 nrg->from = f;
208                 nrg->to   = f;
209                 INIT_LIST_HEAD(&nrg->link);
210                 list_add(&nrg->link, rg->link.prev);
211
212                 return t - f;
213         }
214
215         /* Round our left edge to the current segment if it encloses us. */
216         if (f > rg->from)
217                 f = rg->from;
218         chg = t - f;
219
220         /* Check for and consume any regions we now overlap with. */
221         list_for_each_entry(rg, rg->link.prev, link) {
222                 if (&rg->link == head)
223                         break;
224                 if (rg->from > t)
225                         return chg;
226
227                 /* We overlap with this area, if it extends further than
228                  * us then we must extend ourselves.  Account for its
229                  * existing reservation. */
230                 if (rg->to > t) {
231                         chg += rg->to - t;
232                         t = rg->to;
233                 }
234                 chg -= rg->to - rg->from;
235         }
236         return chg;
237 }
238
239 static long region_truncate(struct list_head *head, long end)
240 {
241         struct file_region *rg, *trg;
242         long chg = 0;
243
244         /* Locate the region we are either in or before. */
245         list_for_each_entry(rg, head, link)
246                 if (end <= rg->to)
247                         break;
248         if (&rg->link == head)
249                 return 0;
250
251         /* If we are in the middle of a region then adjust it. */
252         if (end > rg->from) {
253                 chg = rg->to - end;
254                 rg->to = end;
255                 rg = list_entry(rg->link.next, typeof(*rg), link);
256         }
257
258         /* Drop any remaining regions. */
259         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
260                 if (&rg->link == head)
261                         break;
262                 chg += rg->to - rg->from;
263                 list_del(&rg->link);
264                 kfree(rg);
265         }
266         return chg;
267 }
268
269 static long region_count(struct list_head *head, long f, long t)
270 {
271         struct file_region *rg;
272         long chg = 0;
273
274         /* Locate each segment we overlap with, and count that overlap. */
275         list_for_each_entry(rg, head, link) {
276                 long seg_from;
277                 long seg_to;
278
279                 if (rg->to <= f)
280                         continue;
281                 if (rg->from >= t)
282                         break;
283
284                 seg_from = max(rg->from, f);
285                 seg_to = min(rg->to, t);
286
287                 chg += seg_to - seg_from;
288         }
289
290         return chg;
291 }
292
293 /*
294  * Convert the address within this vma to the page offset within
295  * the mapping, in pagecache page units; huge pages here.
296  */
297 static pgoff_t vma_hugecache_offset(struct hstate *h,
298                         struct vm_area_struct *vma, unsigned long address)
299 {
300         return ((address - vma->vm_start) >> huge_page_shift(h)) +
301                         (vma->vm_pgoff >> huge_page_order(h));
302 }
303
304 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
305                                      unsigned long address)
306 {
307         return vma_hugecache_offset(hstate_vma(vma), vma, address);
308 }
309
310 /*
311  * Return the size of the pages allocated when backing a VMA. In the majority
312  * cases this will be same size as used by the page table entries.
313  */
314 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
315 {
316         struct hstate *hstate;
317
318         if (!is_vm_hugetlb_page(vma))
319                 return PAGE_SIZE;
320
321         hstate = hstate_vma(vma);
322
323         return 1UL << (hstate->order + PAGE_SHIFT);
324 }
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
326
327 /*
328  * Return the page size being used by the MMU to back a VMA. In the majority
329  * of cases, the page size used by the kernel matches the MMU size. On
330  * architectures where it differs, an architecture-specific version of this
331  * function is required.
332  */
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
335 {
336         return vma_kernel_pagesize(vma);
337 }
338 #endif
339
340 /*
341  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
342  * bits of the reservation map pointer, which are always clear due to
343  * alignment.
344  */
345 #define HPAGE_RESV_OWNER    (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
348
349 /*
350  * These helpers are used to track how many pages are reserved for
351  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352  * is guaranteed to have their future faults succeed.
353  *
354  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355  * the reserve counters are updated with the hugetlb_lock held. It is safe
356  * to reset the VMA at fork() time as it is not in use yet and there is no
357  * chance of the global counters getting corrupted as a result of the values.
358  *
359  * The private mapping reservation is represented in a subtly different
360  * manner to a shared mapping.  A shared mapping has a region map associated
361  * with the underlying file, this region map represents the backing file
362  * pages which have ever had a reservation assigned which this persists even
363  * after the page is instantiated.  A private mapping has a region map
364  * associated with the original mmap which is attached to all VMAs which
365  * reference it, this region map represents those offsets which have consumed
366  * reservation ie. where pages have been instantiated.
367  */
368 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
369 {
370         return (unsigned long)vma->vm_private_data;
371 }
372
373 static void set_vma_private_data(struct vm_area_struct *vma,
374                                                         unsigned long value)
375 {
376         vma->vm_private_data = (void *)value;
377 }
378
379 struct resv_map {
380         struct kref refs;
381         struct list_head regions;
382 };
383
384 static struct resv_map *resv_map_alloc(void)
385 {
386         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
387         if (!resv_map)
388                 return NULL;
389
390         kref_init(&resv_map->refs);
391         INIT_LIST_HEAD(&resv_map->regions);
392
393         return resv_map;
394 }
395
396 static void resv_map_release(struct kref *ref)
397 {
398         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
399
400         /* Clear out any active regions before we release the map. */
401         region_truncate(&resv_map->regions, 0);
402         kfree(resv_map);
403 }
404
405 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
406 {
407         VM_BUG_ON(!is_vm_hugetlb_page(vma));
408         if (!(vma->vm_flags & VM_MAYSHARE))
409                 return (struct resv_map *)(get_vma_private_data(vma) &
410                                                         ~HPAGE_RESV_MASK);
411         return NULL;
412 }
413
414 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
415 {
416         VM_BUG_ON(!is_vm_hugetlb_page(vma));
417         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
418
419         set_vma_private_data(vma, (get_vma_private_data(vma) &
420                                 HPAGE_RESV_MASK) | (unsigned long)map);
421 }
422
423 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
424 {
425         VM_BUG_ON(!is_vm_hugetlb_page(vma));
426         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
427
428         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
429 }
430
431 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
432 {
433         VM_BUG_ON(!is_vm_hugetlb_page(vma));
434
435         return (get_vma_private_data(vma) & flag) != 0;
436 }
437
438 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
439 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
440 {
441         VM_BUG_ON(!is_vm_hugetlb_page(vma));
442         if (!(vma->vm_flags & VM_MAYSHARE))
443                 vma->vm_private_data = (void *)0;
444 }
445
446 /* Returns true if the VMA has associated reserve pages */
447 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
448 {
449         if (vma->vm_flags & VM_NORESERVE) {
450                 /*
451                  * This address is already reserved by other process(chg == 0),
452                  * so, we should decrement reserved count. Without decrementing,
453                  * reserve count remains after releasing inode, because this
454                  * allocated page will go into page cache and is regarded as
455                  * coming from reserved pool in releasing step.  Currently, we
456                  * don't have any other solution to deal with this situation
457                  * properly, so add work-around here.
458                  */
459                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
460                         return 1;
461                 else
462                         return 0;
463         }
464
465         /* Shared mappings always use reserves */
466         if (vma->vm_flags & VM_MAYSHARE)
467                 return 1;
468
469         /*
470          * Only the process that called mmap() has reserves for
471          * private mappings.
472          */
473         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
474                 return 1;
475
476         return 0;
477 }
478
479 static void copy_gigantic_page(struct page *dst, struct page *src)
480 {
481         int i;
482         struct hstate *h = page_hstate(src);
483         struct page *dst_base = dst;
484         struct page *src_base = src;
485
486         for (i = 0; i < pages_per_huge_page(h); ) {
487                 cond_resched();
488                 copy_highpage(dst, src);
489
490                 i++;
491                 dst = mem_map_next(dst, dst_base, i);
492                 src = mem_map_next(src, src_base, i);
493         }
494 }
495
496 void copy_huge_page(struct page *dst, struct page *src)
497 {
498         int i;
499         struct hstate *h = page_hstate(src);
500
501         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
502                 copy_gigantic_page(dst, src);
503                 return;
504         }
505
506         might_sleep();
507         for (i = 0; i < pages_per_huge_page(h); i++) {
508                 cond_resched();
509                 copy_highpage(dst + i, src + i);
510         }
511 }
512
513 static void enqueue_huge_page(struct hstate *h, struct page *page)
514 {
515         int nid = page_to_nid(page);
516         list_move(&page->lru, &h->hugepage_freelists[nid]);
517         h->free_huge_pages++;
518         h->free_huge_pages_node[nid]++;
519 }
520
521 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
522 {
523         struct page *page;
524
525         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
526                 if (!is_migrate_isolate_page(page))
527                         break;
528         /*
529          * if 'non-isolated free hugepage' not found on the list,
530          * the allocation fails.
531          */
532         if (&h->hugepage_freelists[nid] == &page->lru)
533                 return NULL;
534         list_move(&page->lru, &h->hugepage_activelist);
535         set_page_refcounted(page);
536         h->free_huge_pages--;
537         h->free_huge_pages_node[nid]--;
538         return page;
539 }
540
541 static struct page *dequeue_huge_page_vma(struct hstate *h,
542                                 struct vm_area_struct *vma,
543                                 unsigned long address, int avoid_reserve,
544                                 long chg)
545 {
546         struct page *page = NULL;
547         struct mempolicy *mpol;
548         nodemask_t *nodemask;
549         struct zonelist *zonelist;
550         struct zone *zone;
551         struct zoneref *z;
552         unsigned int cpuset_mems_cookie;
553
554 retry_cpuset:
555         cpuset_mems_cookie = get_mems_allowed();
556         zonelist = huge_zonelist(vma, address,
557                                         htlb_alloc_mask, &mpol, &nodemask);
558         /*
559          * A child process with MAP_PRIVATE mappings created by their parent
560          * have no page reserves. This check ensures that reservations are
561          * not "stolen". The child may still get SIGKILLed
562          */
563         if (!vma_has_reserves(vma, chg) &&
564                         h->free_huge_pages - h->resv_huge_pages == 0)
565                 goto err;
566
567         /* If reserves cannot be used, ensure enough pages are in the pool */
568         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
569                 goto err;
570
571         for_each_zone_zonelist_nodemask(zone, z, zonelist,
572                                                 MAX_NR_ZONES - 1, nodemask) {
573                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
574                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
575                         if (page) {
576                                 if (avoid_reserve)
577                                         break;
578                                 if (!vma_has_reserves(vma, chg))
579                                         break;
580
581                                 SetPagePrivate(page);
582                                 h->resv_huge_pages--;
583                                 break;
584                         }
585                 }
586         }
587
588         mpol_cond_put(mpol);
589         if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
590                 goto retry_cpuset;
591         return page;
592
593 err:
594         mpol_cond_put(mpol);
595         return NULL;
596 }
597
598 static void update_and_free_page(struct hstate *h, struct page *page)
599 {
600         int i;
601
602         VM_BUG_ON(h->order >= MAX_ORDER);
603
604         h->nr_huge_pages--;
605         h->nr_huge_pages_node[page_to_nid(page)]--;
606         for (i = 0; i < pages_per_huge_page(h); i++) {
607                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
608                                 1 << PG_referenced | 1 << PG_dirty |
609                                 1 << PG_active | 1 << PG_reserved |
610                                 1 << PG_private | 1 << PG_writeback);
611         }
612         VM_BUG_ON(hugetlb_cgroup_from_page(page));
613         set_compound_page_dtor(page, NULL);
614         set_page_refcounted(page);
615         arch_release_hugepage(page);
616         __free_pages(page, huge_page_order(h));
617 }
618
619 struct hstate *size_to_hstate(unsigned long size)
620 {
621         struct hstate *h;
622
623         for_each_hstate(h) {
624                 if (huge_page_size(h) == size)
625                         return h;
626         }
627         return NULL;
628 }
629
630 static void free_huge_page(struct page *page)
631 {
632         /*
633          * Can't pass hstate in here because it is called from the
634          * compound page destructor.
635          */
636         struct hstate *h = page_hstate(page);
637         int nid = page_to_nid(page);
638         struct hugepage_subpool *spool =
639                 (struct hugepage_subpool *)page_private(page);
640         bool restore_reserve;
641
642         set_page_private(page, 0);
643         page->mapping = NULL;
644         BUG_ON(page_count(page));
645         BUG_ON(page_mapcount(page));
646         restore_reserve = PagePrivate(page);
647
648         spin_lock(&hugetlb_lock);
649         hugetlb_cgroup_uncharge_page(hstate_index(h),
650                                      pages_per_huge_page(h), page);
651         if (restore_reserve)
652                 h->resv_huge_pages++;
653
654         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
655                 /* remove the page from active list */
656                 list_del(&page->lru);
657                 update_and_free_page(h, page);
658                 h->surplus_huge_pages--;
659                 h->surplus_huge_pages_node[nid]--;
660         } else {
661                 arch_clear_hugepage_flags(page);
662                 enqueue_huge_page(h, page);
663         }
664         spin_unlock(&hugetlb_lock);
665         hugepage_subpool_put_pages(spool, 1);
666 }
667
668 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
669 {
670         INIT_LIST_HEAD(&page->lru);
671         set_compound_page_dtor(page, free_huge_page);
672         spin_lock(&hugetlb_lock);
673         set_hugetlb_cgroup(page, NULL);
674         h->nr_huge_pages++;
675         h->nr_huge_pages_node[nid]++;
676         spin_unlock(&hugetlb_lock);
677         put_page(page); /* free it into the hugepage allocator */
678 }
679
680 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
681 {
682         int i;
683         int nr_pages = 1 << order;
684         struct page *p = page + 1;
685
686         /* we rely on prep_new_huge_page to set the destructor */
687         set_compound_order(page, order);
688         __SetPageHead(page);
689         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
690                 __SetPageTail(p);
691                 set_page_count(p, 0);
692                 p->first_page = page;
693         }
694 }
695
696 /*
697  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
698  * transparent huge pages.  See the PageTransHuge() documentation for more
699  * details.
700  */
701 int PageHuge(struct page *page)
702 {
703         compound_page_dtor *dtor;
704
705         if (!PageCompound(page))
706                 return 0;
707
708         page = compound_head(page);
709         dtor = get_compound_page_dtor(page);
710
711         return dtor == free_huge_page;
712 }
713 EXPORT_SYMBOL_GPL(PageHuge);
714
715 /*
716  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
717  * normal or transparent huge pages.
718  */
719 int PageHeadHuge(struct page *page_head)
720 {
721         compound_page_dtor *dtor;
722
723         if (!PageHead(page_head))
724                 return 0;
725
726         dtor = get_compound_page_dtor(page_head);
727
728         return dtor == free_huge_page;
729 }
730 EXPORT_SYMBOL_GPL(PageHeadHuge);
731
732 pgoff_t __basepage_index(struct page *page)
733 {
734         struct page *page_head = compound_head(page);
735         pgoff_t index = page_index(page_head);
736         unsigned long compound_idx;
737
738         if (!PageHuge(page_head))
739                 return page_index(page);
740
741         if (compound_order(page_head) >= MAX_ORDER)
742                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
743         else
744                 compound_idx = page - page_head;
745
746         return (index << compound_order(page_head)) + compound_idx;
747 }
748
749 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
750 {
751         struct page *page;
752
753         if (h->order >= MAX_ORDER)
754                 return NULL;
755
756         page = alloc_pages_exact_node(nid,
757                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
758                                                 __GFP_REPEAT|__GFP_NOWARN,
759                 huge_page_order(h));
760         if (page) {
761                 if (arch_prepare_hugepage(page)) {
762                         __free_pages(page, huge_page_order(h));
763                         return NULL;
764                 }
765                 prep_new_huge_page(h, page, nid);
766         }
767
768         return page;
769 }
770
771 /*
772  * common helper functions for hstate_next_node_to_{alloc|free}.
773  * We may have allocated or freed a huge page based on a different
774  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
775  * be outside of *nodes_allowed.  Ensure that we use an allowed
776  * node for alloc or free.
777  */
778 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
779 {
780         nid = next_node(nid, *nodes_allowed);
781         if (nid == MAX_NUMNODES)
782                 nid = first_node(*nodes_allowed);
783         VM_BUG_ON(nid >= MAX_NUMNODES);
784
785         return nid;
786 }
787
788 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
789 {
790         if (!node_isset(nid, *nodes_allowed))
791                 nid = next_node_allowed(nid, nodes_allowed);
792         return nid;
793 }
794
795 /*
796  * returns the previously saved node ["this node"] from which to
797  * allocate a persistent huge page for the pool and advance the
798  * next node from which to allocate, handling wrap at end of node
799  * mask.
800  */
801 static int hstate_next_node_to_alloc(struct hstate *h,
802                                         nodemask_t *nodes_allowed)
803 {
804         int nid;
805
806         VM_BUG_ON(!nodes_allowed);
807
808         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
809         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
810
811         return nid;
812 }
813
814 /*
815  * helper for free_pool_huge_page() - return the previously saved
816  * node ["this node"] from which to free a huge page.  Advance the
817  * next node id whether or not we find a free huge page to free so
818  * that the next attempt to free addresses the next node.
819  */
820 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
821 {
822         int nid;
823
824         VM_BUG_ON(!nodes_allowed);
825
826         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
827         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
828
829         return nid;
830 }
831
832 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
833         for (nr_nodes = nodes_weight(*mask);                            \
834                 nr_nodes > 0 &&                                         \
835                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
836                 nr_nodes--)
837
838 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
839         for (nr_nodes = nodes_weight(*mask);                            \
840                 nr_nodes > 0 &&                                         \
841                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
842                 nr_nodes--)
843
844 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
845 {
846         struct page *page;
847         int nr_nodes, node;
848         int ret = 0;
849
850         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
851                 page = alloc_fresh_huge_page_node(h, node);
852                 if (page) {
853                         ret = 1;
854                         break;
855                 }
856         }
857
858         if (ret)
859                 count_vm_event(HTLB_BUDDY_PGALLOC);
860         else
861                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
862
863         return ret;
864 }
865
866 /*
867  * Free huge page from pool from next node to free.
868  * Attempt to keep persistent huge pages more or less
869  * balanced over allowed nodes.
870  * Called with hugetlb_lock locked.
871  */
872 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
873                                                          bool acct_surplus)
874 {
875         int nr_nodes, node;
876         int ret = 0;
877
878         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
879                 /*
880                  * If we're returning unused surplus pages, only examine
881                  * nodes with surplus pages.
882                  */
883                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
884                     !list_empty(&h->hugepage_freelists[node])) {
885                         struct page *page =
886                                 list_entry(h->hugepage_freelists[node].next,
887                                           struct page, lru);
888                         list_del(&page->lru);
889                         h->free_huge_pages--;
890                         h->free_huge_pages_node[node]--;
891                         if (acct_surplus) {
892                                 h->surplus_huge_pages--;
893                                 h->surplus_huge_pages_node[node]--;
894                         }
895                         update_and_free_page(h, page);
896                         ret = 1;
897                         break;
898                 }
899         }
900
901         return ret;
902 }
903
904 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
905 {
906         struct page *page;
907         unsigned int r_nid;
908
909         if (h->order >= MAX_ORDER)
910                 return NULL;
911
912         /*
913          * Assume we will successfully allocate the surplus page to
914          * prevent racing processes from causing the surplus to exceed
915          * overcommit
916          *
917          * This however introduces a different race, where a process B
918          * tries to grow the static hugepage pool while alloc_pages() is
919          * called by process A. B will only examine the per-node
920          * counters in determining if surplus huge pages can be
921          * converted to normal huge pages in adjust_pool_surplus(). A
922          * won't be able to increment the per-node counter, until the
923          * lock is dropped by B, but B doesn't drop hugetlb_lock until
924          * no more huge pages can be converted from surplus to normal
925          * state (and doesn't try to convert again). Thus, we have a
926          * case where a surplus huge page exists, the pool is grown, and
927          * the surplus huge page still exists after, even though it
928          * should just have been converted to a normal huge page. This
929          * does not leak memory, though, as the hugepage will be freed
930          * once it is out of use. It also does not allow the counters to
931          * go out of whack in adjust_pool_surplus() as we don't modify
932          * the node values until we've gotten the hugepage and only the
933          * per-node value is checked there.
934          */
935         spin_lock(&hugetlb_lock);
936         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
937                 spin_unlock(&hugetlb_lock);
938                 return NULL;
939         } else {
940                 h->nr_huge_pages++;
941                 h->surplus_huge_pages++;
942         }
943         spin_unlock(&hugetlb_lock);
944
945         if (nid == NUMA_NO_NODE)
946                 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
947                                    __GFP_REPEAT|__GFP_NOWARN,
948                                    huge_page_order(h));
949         else
950                 page = alloc_pages_exact_node(nid,
951                         htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
952                         __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
953
954         if (page && arch_prepare_hugepage(page)) {
955                 __free_pages(page, huge_page_order(h));
956                 page = NULL;
957         }
958
959         spin_lock(&hugetlb_lock);
960         if (page) {
961                 INIT_LIST_HEAD(&page->lru);
962                 r_nid = page_to_nid(page);
963                 set_compound_page_dtor(page, free_huge_page);
964                 set_hugetlb_cgroup(page, NULL);
965                 /*
966                  * We incremented the global counters already
967                  */
968                 h->nr_huge_pages_node[r_nid]++;
969                 h->surplus_huge_pages_node[r_nid]++;
970                 __count_vm_event(HTLB_BUDDY_PGALLOC);
971         } else {
972                 h->nr_huge_pages--;
973                 h->surplus_huge_pages--;
974                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
975         }
976         spin_unlock(&hugetlb_lock);
977
978         return page;
979 }
980
981 /*
982  * This allocation function is useful in the context where vma is irrelevant.
983  * E.g. soft-offlining uses this function because it only cares physical
984  * address of error page.
985  */
986 struct page *alloc_huge_page_node(struct hstate *h, int nid)
987 {
988         struct page *page = NULL;
989
990         spin_lock(&hugetlb_lock);
991         if (h->free_huge_pages - h->resv_huge_pages > 0)
992                 page = dequeue_huge_page_node(h, nid);
993         spin_unlock(&hugetlb_lock);
994
995         if (!page)
996                 page = alloc_buddy_huge_page(h, nid);
997
998         return page;
999 }
1000
1001 /*
1002  * Increase the hugetlb pool such that it can accommodate a reservation
1003  * of size 'delta'.
1004  */
1005 static int gather_surplus_pages(struct hstate *h, int delta)
1006 {
1007         struct list_head surplus_list;
1008         struct page *page, *tmp;
1009         int ret, i;
1010         int needed, allocated;
1011         bool alloc_ok = true;
1012
1013         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1014         if (needed <= 0) {
1015                 h->resv_huge_pages += delta;
1016                 return 0;
1017         }
1018
1019         allocated = 0;
1020         INIT_LIST_HEAD(&surplus_list);
1021
1022         ret = -ENOMEM;
1023 retry:
1024         spin_unlock(&hugetlb_lock);
1025         for (i = 0; i < needed; i++) {
1026                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1027                 if (!page) {
1028                         alloc_ok = false;
1029                         break;
1030                 }
1031                 list_add(&page->lru, &surplus_list);
1032         }
1033         allocated += i;
1034
1035         /*
1036          * After retaking hugetlb_lock, we need to recalculate 'needed'
1037          * because either resv_huge_pages or free_huge_pages may have changed.
1038          */
1039         spin_lock(&hugetlb_lock);
1040         needed = (h->resv_huge_pages + delta) -
1041                         (h->free_huge_pages + allocated);
1042         if (needed > 0) {
1043                 if (alloc_ok)
1044                         goto retry;
1045                 /*
1046                  * We were not able to allocate enough pages to
1047                  * satisfy the entire reservation so we free what
1048                  * we've allocated so far.
1049                  */
1050                 goto free;
1051         }
1052         /*
1053          * The surplus_list now contains _at_least_ the number of extra pages
1054          * needed to accommodate the reservation.  Add the appropriate number
1055          * of pages to the hugetlb pool and free the extras back to the buddy
1056          * allocator.  Commit the entire reservation here to prevent another
1057          * process from stealing the pages as they are added to the pool but
1058          * before they are reserved.
1059          */
1060         needed += allocated;
1061         h->resv_huge_pages += delta;
1062         ret = 0;
1063
1064         /* Free the needed pages to the hugetlb pool */
1065         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1066                 if ((--needed) < 0)
1067                         break;
1068                 /*
1069                  * This page is now managed by the hugetlb allocator and has
1070                  * no users -- drop the buddy allocator's reference.
1071                  */
1072                 put_page_testzero(page);
1073                 VM_BUG_ON(page_count(page));
1074                 enqueue_huge_page(h, page);
1075         }
1076 free:
1077         spin_unlock(&hugetlb_lock);
1078
1079         /* Free unnecessary surplus pages to the buddy allocator */
1080         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1081                 put_page(page);
1082         spin_lock(&hugetlb_lock);
1083
1084         return ret;
1085 }
1086
1087 /*
1088  * When releasing a hugetlb pool reservation, any surplus pages that were
1089  * allocated to satisfy the reservation must be explicitly freed if they were
1090  * never used.
1091  * Called with hugetlb_lock held.
1092  */
1093 static void return_unused_surplus_pages(struct hstate *h,
1094                                         unsigned long unused_resv_pages)
1095 {
1096         unsigned long nr_pages;
1097
1098         /* Uncommit the reservation */
1099         h->resv_huge_pages -= unused_resv_pages;
1100
1101         /* Cannot return gigantic pages currently */
1102         if (h->order >= MAX_ORDER)
1103                 return;
1104
1105         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1106
1107         /*
1108          * We want to release as many surplus pages as possible, spread
1109          * evenly across all nodes with memory. Iterate across these nodes
1110          * until we can no longer free unreserved surplus pages. This occurs
1111          * when the nodes with surplus pages have no free pages.
1112          * free_pool_huge_page() will balance the the freed pages across the
1113          * on-line nodes with memory and will handle the hstate accounting.
1114          */
1115         while (nr_pages--) {
1116                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1117                         break;
1118                 cond_resched_lock(&hugetlb_lock);
1119         }
1120 }
1121
1122 /*
1123  * Determine if the huge page at addr within the vma has an associated
1124  * reservation.  Where it does not we will need to logically increase
1125  * reservation and actually increase subpool usage before an allocation
1126  * can occur.  Where any new reservation would be required the
1127  * reservation change is prepared, but not committed.  Once the page
1128  * has been allocated from the subpool and instantiated the change should
1129  * be committed via vma_commit_reservation.  No action is required on
1130  * failure.
1131  */
1132 static long vma_needs_reservation(struct hstate *h,
1133                         struct vm_area_struct *vma, unsigned long addr)
1134 {
1135         struct address_space *mapping = vma->vm_file->f_mapping;
1136         struct inode *inode = mapping->host;
1137
1138         if (vma->vm_flags & VM_MAYSHARE) {
1139                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1140                 return region_chg(&inode->i_mapping->private_list,
1141                                                         idx, idx + 1);
1142
1143         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1144                 return 1;
1145
1146         } else  {
1147                 long err;
1148                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1149                 struct resv_map *resv = vma_resv_map(vma);
1150
1151                 err = region_chg(&resv->regions, idx, idx + 1);
1152                 if (err < 0)
1153                         return err;
1154                 return 0;
1155         }
1156 }
1157 static void vma_commit_reservation(struct hstate *h,
1158                         struct vm_area_struct *vma, unsigned long addr)
1159 {
1160         struct address_space *mapping = vma->vm_file->f_mapping;
1161         struct inode *inode = mapping->host;
1162
1163         if (vma->vm_flags & VM_MAYSHARE) {
1164                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1165                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1166
1167         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1168                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1169                 struct resv_map *resv = vma_resv_map(vma);
1170
1171                 /* Mark this page used in the map. */
1172                 region_add(&resv->regions, idx, idx + 1);
1173         }
1174 }
1175
1176 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1177                                     unsigned long addr, int avoid_reserve)
1178 {
1179         struct hugepage_subpool *spool = subpool_vma(vma);
1180         struct hstate *h = hstate_vma(vma);
1181         struct page *page;
1182         long chg;
1183         int ret, idx;
1184         struct hugetlb_cgroup *h_cg;
1185
1186         idx = hstate_index(h);
1187         /*
1188          * Processes that did not create the mapping will have no
1189          * reserves and will not have accounted against subpool
1190          * limit. Check that the subpool limit can be made before
1191          * satisfying the allocation MAP_NORESERVE mappings may also
1192          * need pages and subpool limit allocated allocated if no reserve
1193          * mapping overlaps.
1194          */
1195         chg = vma_needs_reservation(h, vma, addr);
1196         if (chg < 0)
1197                 return ERR_PTR(-ENOMEM);
1198         if (chg || avoid_reserve)
1199                 if (hugepage_subpool_get_pages(spool, 1))
1200                         return ERR_PTR(-ENOSPC);
1201
1202         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1203         if (ret) {
1204                 if (chg || avoid_reserve)
1205                         hugepage_subpool_put_pages(spool, 1);
1206                 return ERR_PTR(-ENOSPC);
1207         }
1208         spin_lock(&hugetlb_lock);
1209         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1210         if (!page) {
1211                 spin_unlock(&hugetlb_lock);
1212                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1213                 if (!page) {
1214                         hugetlb_cgroup_uncharge_cgroup(idx,
1215                                                        pages_per_huge_page(h),
1216                                                        h_cg);
1217                         if (chg || avoid_reserve)
1218                                 hugepage_subpool_put_pages(spool, 1);
1219                         return ERR_PTR(-ENOSPC);
1220                 }
1221                 spin_lock(&hugetlb_lock);
1222                 list_move(&page->lru, &h->hugepage_activelist);
1223                 /* Fall through */
1224         }
1225         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1226         spin_unlock(&hugetlb_lock);
1227
1228         set_page_private(page, (unsigned long)spool);
1229
1230         vma_commit_reservation(h, vma, addr);
1231         return page;
1232 }
1233
1234 int __weak alloc_bootmem_huge_page(struct hstate *h)
1235 {
1236         struct huge_bootmem_page *m;
1237         int nr_nodes, node;
1238
1239         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1240                 void *addr;
1241
1242                 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
1243                                 huge_page_size(h), huge_page_size(h), 0);
1244
1245                 if (addr) {
1246                         /*
1247                          * Use the beginning of the huge page to store the
1248                          * huge_bootmem_page struct (until gather_bootmem
1249                          * puts them into the mem_map).
1250                          */
1251                         m = addr;
1252                         goto found;
1253                 }
1254         }
1255         return 0;
1256
1257 found:
1258         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1259         /* Put them into a private list first because mem_map is not up yet */
1260         list_add(&m->list, &huge_boot_pages);
1261         m->hstate = h;
1262         return 1;
1263 }
1264
1265 static void prep_compound_huge_page(struct page *page, int order)
1266 {
1267         if (unlikely(order > (MAX_ORDER - 1)))
1268                 prep_compound_gigantic_page(page, order);
1269         else
1270                 prep_compound_page(page, order);
1271 }
1272
1273 /* Put bootmem huge pages into the standard lists after mem_map is up */
1274 static void __init gather_bootmem_prealloc(void)
1275 {
1276         struct huge_bootmem_page *m;
1277
1278         list_for_each_entry(m, &huge_boot_pages, list) {
1279                 struct hstate *h = m->hstate;
1280                 struct page *page;
1281
1282 #ifdef CONFIG_HIGHMEM
1283                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1284                 free_bootmem_late((unsigned long)m,
1285                                   sizeof(struct huge_bootmem_page));
1286 #else
1287                 page = virt_to_page(m);
1288 #endif
1289                 __ClearPageReserved(page);
1290                 WARN_ON(page_count(page) != 1);
1291                 prep_compound_huge_page(page, h->order);
1292                 prep_new_huge_page(h, page, page_to_nid(page));
1293                 /*
1294                  * If we had gigantic hugepages allocated at boot time, we need
1295                  * to restore the 'stolen' pages to totalram_pages in order to
1296                  * fix confusing memory reports from free(1) and another
1297                  * side-effects, like CommitLimit going negative.
1298                  */
1299                 if (h->order > (MAX_ORDER - 1))
1300                         totalram_pages += 1 << h->order;
1301         }
1302 }
1303
1304 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1305 {
1306         unsigned long i;
1307
1308         for (i = 0; i < h->max_huge_pages; ++i) {
1309                 if (h->order >= MAX_ORDER) {
1310                         if (!alloc_bootmem_huge_page(h))
1311                                 break;
1312                 } else if (!alloc_fresh_huge_page(h,
1313                                          &node_states[N_MEMORY]))
1314                         break;
1315         }
1316         h->max_huge_pages = i;
1317 }
1318
1319 static void __init hugetlb_init_hstates(void)
1320 {
1321         struct hstate *h;
1322
1323         for_each_hstate(h) {
1324                 /* oversize hugepages were init'ed in early boot */
1325                 if (h->order < MAX_ORDER)
1326                         hugetlb_hstate_alloc_pages(h);
1327         }
1328 }
1329
1330 static char * __init memfmt(char *buf, unsigned long n)
1331 {
1332         if (n >= (1UL << 30))
1333                 sprintf(buf, "%lu GB", n >> 30);
1334         else if (n >= (1UL << 20))
1335                 sprintf(buf, "%lu MB", n >> 20);
1336         else
1337                 sprintf(buf, "%lu KB", n >> 10);
1338         return buf;
1339 }
1340
1341 static void __init report_hugepages(void)
1342 {
1343         struct hstate *h;
1344
1345         for_each_hstate(h) {
1346                 char buf[32];
1347                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1348                         memfmt(buf, huge_page_size(h)),
1349                         h->free_huge_pages);
1350         }
1351 }
1352
1353 #ifdef CONFIG_HIGHMEM
1354 static void try_to_free_low(struct hstate *h, unsigned long count,
1355                                                 nodemask_t *nodes_allowed)
1356 {
1357         int i;
1358
1359         if (h->order >= MAX_ORDER)
1360                 return;
1361
1362         for_each_node_mask(i, *nodes_allowed) {
1363                 struct page *page, *next;
1364                 struct list_head *freel = &h->hugepage_freelists[i];
1365                 list_for_each_entry_safe(page, next, freel, lru) {
1366                         if (count >= h->nr_huge_pages)
1367                                 return;
1368                         if (PageHighMem(page))
1369                                 continue;
1370                         list_del(&page->lru);
1371                         update_and_free_page(h, page);
1372                         h->free_huge_pages--;
1373                         h->free_huge_pages_node[page_to_nid(page)]--;
1374                 }
1375         }
1376 }
1377 #else
1378 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1379                                                 nodemask_t *nodes_allowed)
1380 {
1381 }
1382 #endif
1383
1384 /*
1385  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1386  * balanced by operating on them in a round-robin fashion.
1387  * Returns 1 if an adjustment was made.
1388  */
1389 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1390                                 int delta)
1391 {
1392         int nr_nodes, node;
1393
1394         VM_BUG_ON(delta != -1 && delta != 1);
1395
1396         if (delta < 0) {
1397                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1398                         if (h->surplus_huge_pages_node[node])
1399                                 goto found;
1400                 }
1401         } else {
1402                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1403                         if (h->surplus_huge_pages_node[node] <
1404                                         h->nr_huge_pages_node[node])
1405                                 goto found;
1406                 }
1407         }
1408         return 0;
1409
1410 found:
1411         h->surplus_huge_pages += delta;
1412         h->surplus_huge_pages_node[node] += delta;
1413         return 1;
1414 }
1415
1416 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1417 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1418                                                 nodemask_t *nodes_allowed)
1419 {
1420         unsigned long min_count, ret;
1421
1422         if (h->order >= MAX_ORDER)
1423                 return h->max_huge_pages;
1424
1425         /*
1426          * Increase the pool size
1427          * First take pages out of surplus state.  Then make up the
1428          * remaining difference by allocating fresh huge pages.
1429          *
1430          * We might race with alloc_buddy_huge_page() here and be unable
1431          * to convert a surplus huge page to a normal huge page. That is
1432          * not critical, though, it just means the overall size of the
1433          * pool might be one hugepage larger than it needs to be, but
1434          * within all the constraints specified by the sysctls.
1435          */
1436         spin_lock(&hugetlb_lock);
1437         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1438                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1439                         break;
1440         }
1441
1442         while (count > persistent_huge_pages(h)) {
1443                 /*
1444                  * If this allocation races such that we no longer need the
1445                  * page, free_huge_page will handle it by freeing the page
1446                  * and reducing the surplus.
1447                  */
1448                 spin_unlock(&hugetlb_lock);
1449                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1450                 spin_lock(&hugetlb_lock);
1451                 if (!ret)
1452                         goto out;
1453
1454                 /* Bail for signals. Probably ctrl-c from user */
1455                 if (signal_pending(current))
1456                         goto out;
1457         }
1458
1459         /*
1460          * Decrease the pool size
1461          * First return free pages to the buddy allocator (being careful
1462          * to keep enough around to satisfy reservations).  Then place
1463          * pages into surplus state as needed so the pool will shrink
1464          * to the desired size as pages become free.
1465          *
1466          * By placing pages into the surplus state independent of the
1467          * overcommit value, we are allowing the surplus pool size to
1468          * exceed overcommit. There are few sane options here. Since
1469          * alloc_buddy_huge_page() is checking the global counter,
1470          * though, we'll note that we're not allowed to exceed surplus
1471          * and won't grow the pool anywhere else. Not until one of the
1472          * sysctls are changed, or the surplus pages go out of use.
1473          */
1474         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1475         min_count = max(count, min_count);
1476         try_to_free_low(h, min_count, nodes_allowed);
1477         while (min_count < persistent_huge_pages(h)) {
1478                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1479                         break;
1480                 cond_resched_lock(&hugetlb_lock);
1481         }
1482         while (count < persistent_huge_pages(h)) {
1483                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1484                         break;
1485         }
1486 out:
1487         ret = persistent_huge_pages(h);
1488         spin_unlock(&hugetlb_lock);
1489         return ret;
1490 }
1491
1492 #define HSTATE_ATTR_RO(_name) \
1493         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1494
1495 #define HSTATE_ATTR(_name) \
1496         static struct kobj_attribute _name##_attr = \
1497                 __ATTR(_name, 0644, _name##_show, _name##_store)
1498
1499 static struct kobject *hugepages_kobj;
1500 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1501
1502 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1503
1504 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1505 {
1506         int i;
1507
1508         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1509                 if (hstate_kobjs[i] == kobj) {
1510                         if (nidp)
1511                                 *nidp = NUMA_NO_NODE;
1512                         return &hstates[i];
1513                 }
1514
1515         return kobj_to_node_hstate(kobj, nidp);
1516 }
1517
1518 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1519                                         struct kobj_attribute *attr, char *buf)
1520 {
1521         struct hstate *h;
1522         unsigned long nr_huge_pages;
1523         int nid;
1524
1525         h = kobj_to_hstate(kobj, &nid);
1526         if (nid == NUMA_NO_NODE)
1527                 nr_huge_pages = h->nr_huge_pages;
1528         else
1529                 nr_huge_pages = h->nr_huge_pages_node[nid];
1530
1531         return sprintf(buf, "%lu\n", nr_huge_pages);
1532 }
1533
1534 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1535                         struct kobject *kobj, struct kobj_attribute *attr,
1536                         const char *buf, size_t len)
1537 {
1538         int err;
1539         int nid;
1540         unsigned long count;
1541         struct hstate *h;
1542         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1543
1544         err = strict_strtoul(buf, 10, &count);
1545         if (err)
1546                 goto out;
1547
1548         h = kobj_to_hstate(kobj, &nid);
1549         if (h->order >= MAX_ORDER) {
1550                 err = -EINVAL;
1551                 goto out;
1552         }
1553
1554         if (nid == NUMA_NO_NODE) {
1555                 /*
1556                  * global hstate attribute
1557                  */
1558                 if (!(obey_mempolicy &&
1559                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1560                         NODEMASK_FREE(nodes_allowed);
1561                         nodes_allowed = &node_states[N_MEMORY];
1562                 }
1563         } else if (nodes_allowed) {
1564                 /*
1565                  * per node hstate attribute: adjust count to global,
1566                  * but restrict alloc/free to the specified node.
1567                  */
1568                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1569                 init_nodemask_of_node(nodes_allowed, nid);
1570         } else
1571                 nodes_allowed = &node_states[N_MEMORY];
1572
1573         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1574
1575         if (nodes_allowed != &node_states[N_MEMORY])
1576                 NODEMASK_FREE(nodes_allowed);
1577
1578         return len;
1579 out:
1580         NODEMASK_FREE(nodes_allowed);
1581         return err;
1582 }
1583
1584 static ssize_t nr_hugepages_show(struct kobject *kobj,
1585                                        struct kobj_attribute *attr, char *buf)
1586 {
1587         return nr_hugepages_show_common(kobj, attr, buf);
1588 }
1589
1590 static ssize_t nr_hugepages_store(struct kobject *kobj,
1591                struct kobj_attribute *attr, const char *buf, size_t len)
1592 {
1593         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1594 }
1595 HSTATE_ATTR(nr_hugepages);
1596
1597 #ifdef CONFIG_NUMA
1598
1599 /*
1600  * hstate attribute for optionally mempolicy-based constraint on persistent
1601  * huge page alloc/free.
1602  */
1603 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1604                                        struct kobj_attribute *attr, char *buf)
1605 {
1606         return nr_hugepages_show_common(kobj, attr, buf);
1607 }
1608
1609 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1610                struct kobj_attribute *attr, const char *buf, size_t len)
1611 {
1612         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1613 }
1614 HSTATE_ATTR(nr_hugepages_mempolicy);
1615 #endif
1616
1617
1618 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1619                                         struct kobj_attribute *attr, char *buf)
1620 {
1621         struct hstate *h = kobj_to_hstate(kobj, NULL);
1622         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1623 }
1624
1625 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1626                 struct kobj_attribute *attr, const char *buf, size_t count)
1627 {
1628         int err;
1629         unsigned long input;
1630         struct hstate *h = kobj_to_hstate(kobj, NULL);
1631
1632         if (h->order >= MAX_ORDER)
1633                 return -EINVAL;
1634
1635         err = strict_strtoul(buf, 10, &input);
1636         if (err)
1637                 return err;
1638
1639         spin_lock(&hugetlb_lock);
1640         h->nr_overcommit_huge_pages = input;
1641         spin_unlock(&hugetlb_lock);
1642
1643         return count;
1644 }
1645 HSTATE_ATTR(nr_overcommit_hugepages);
1646
1647 static ssize_t free_hugepages_show(struct kobject *kobj,
1648                                         struct kobj_attribute *attr, char *buf)
1649 {
1650         struct hstate *h;
1651         unsigned long free_huge_pages;
1652         int nid;
1653
1654         h = kobj_to_hstate(kobj, &nid);
1655         if (nid == NUMA_NO_NODE)
1656                 free_huge_pages = h->free_huge_pages;
1657         else
1658                 free_huge_pages = h->free_huge_pages_node[nid];
1659
1660         return sprintf(buf, "%lu\n", free_huge_pages);
1661 }
1662 HSTATE_ATTR_RO(free_hugepages);
1663
1664 static ssize_t resv_hugepages_show(struct kobject *kobj,
1665                                         struct kobj_attribute *attr, char *buf)
1666 {
1667         struct hstate *h = kobj_to_hstate(kobj, NULL);
1668         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1669 }
1670 HSTATE_ATTR_RO(resv_hugepages);
1671
1672 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1673                                         struct kobj_attribute *attr, char *buf)
1674 {
1675         struct hstate *h;
1676         unsigned long surplus_huge_pages;
1677         int nid;
1678
1679         h = kobj_to_hstate(kobj, &nid);
1680         if (nid == NUMA_NO_NODE)
1681                 surplus_huge_pages = h->surplus_huge_pages;
1682         else
1683                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1684
1685         return sprintf(buf, "%lu\n", surplus_huge_pages);
1686 }
1687 HSTATE_ATTR_RO(surplus_hugepages);
1688
1689 static struct attribute *hstate_attrs[] = {
1690         &nr_hugepages_attr.attr,
1691         &nr_overcommit_hugepages_attr.attr,
1692         &free_hugepages_attr.attr,
1693         &resv_hugepages_attr.attr,
1694         &surplus_hugepages_attr.attr,
1695 #ifdef CONFIG_NUMA
1696         &nr_hugepages_mempolicy_attr.attr,
1697 #endif
1698         NULL,
1699 };
1700
1701 static struct attribute_group hstate_attr_group = {
1702         .attrs = hstate_attrs,
1703 };
1704
1705 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1706                                     struct kobject **hstate_kobjs,
1707                                     struct attribute_group *hstate_attr_group)
1708 {
1709         int retval;
1710         int hi = hstate_index(h);
1711
1712         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1713         if (!hstate_kobjs[hi])
1714                 return -ENOMEM;
1715
1716         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1717         if (retval)
1718                 kobject_put(hstate_kobjs[hi]);
1719
1720         return retval;
1721 }
1722
1723 static void __init hugetlb_sysfs_init(void)
1724 {
1725         struct hstate *h;
1726         int err;
1727
1728         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1729         if (!hugepages_kobj)
1730                 return;
1731
1732         for_each_hstate(h) {
1733                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1734                                          hstate_kobjs, &hstate_attr_group);
1735                 if (err)
1736                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
1737         }
1738 }
1739
1740 #ifdef CONFIG_NUMA
1741
1742 /*
1743  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1744  * with node devices in node_devices[] using a parallel array.  The array
1745  * index of a node device or _hstate == node id.
1746  * This is here to avoid any static dependency of the node device driver, in
1747  * the base kernel, on the hugetlb module.
1748  */
1749 struct node_hstate {
1750         struct kobject          *hugepages_kobj;
1751         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1752 };
1753 struct node_hstate node_hstates[MAX_NUMNODES];
1754
1755 /*
1756  * A subset of global hstate attributes for node devices
1757  */
1758 static struct attribute *per_node_hstate_attrs[] = {
1759         &nr_hugepages_attr.attr,
1760         &free_hugepages_attr.attr,
1761         &surplus_hugepages_attr.attr,
1762         NULL,
1763 };
1764
1765 static struct attribute_group per_node_hstate_attr_group = {
1766         .attrs = per_node_hstate_attrs,
1767 };
1768
1769 /*
1770  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1771  * Returns node id via non-NULL nidp.
1772  */
1773 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1774 {
1775         int nid;
1776
1777         for (nid = 0; nid < nr_node_ids; nid++) {
1778                 struct node_hstate *nhs = &node_hstates[nid];
1779                 int i;
1780                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1781                         if (nhs->hstate_kobjs[i] == kobj) {
1782                                 if (nidp)
1783                                         *nidp = nid;
1784                                 return &hstates[i];
1785                         }
1786         }
1787
1788         BUG();
1789         return NULL;
1790 }
1791
1792 /*
1793  * Unregister hstate attributes from a single node device.
1794  * No-op if no hstate attributes attached.
1795  */
1796 static void hugetlb_unregister_node(struct node *node)
1797 {
1798         struct hstate *h;
1799         struct node_hstate *nhs = &node_hstates[node->dev.id];
1800
1801         if (!nhs->hugepages_kobj)
1802                 return;         /* no hstate attributes */
1803
1804         for_each_hstate(h) {
1805                 int idx = hstate_index(h);
1806                 if (nhs->hstate_kobjs[idx]) {
1807                         kobject_put(nhs->hstate_kobjs[idx]);
1808                         nhs->hstate_kobjs[idx] = NULL;
1809                 }
1810         }
1811
1812         kobject_put(nhs->hugepages_kobj);
1813         nhs->hugepages_kobj = NULL;
1814 }
1815
1816 /*
1817  * hugetlb module exit:  unregister hstate attributes from node devices
1818  * that have them.
1819  */
1820 static void hugetlb_unregister_all_nodes(void)
1821 {
1822         int nid;
1823
1824         /*
1825          * disable node device registrations.
1826          */
1827         register_hugetlbfs_with_node(NULL, NULL);
1828
1829         /*
1830          * remove hstate attributes from any nodes that have them.
1831          */
1832         for (nid = 0; nid < nr_node_ids; nid++)
1833                 hugetlb_unregister_node(node_devices[nid]);
1834 }
1835
1836 /*
1837  * Register hstate attributes for a single node device.
1838  * No-op if attributes already registered.
1839  */
1840 static void hugetlb_register_node(struct node *node)
1841 {
1842         struct hstate *h;
1843         struct node_hstate *nhs = &node_hstates[node->dev.id];
1844         int err;
1845
1846         if (nhs->hugepages_kobj)
1847                 return;         /* already allocated */
1848
1849         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1850                                                         &node->dev.kobj);
1851         if (!nhs->hugepages_kobj)
1852                 return;
1853
1854         for_each_hstate(h) {
1855                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1856                                                 nhs->hstate_kobjs,
1857                                                 &per_node_hstate_attr_group);
1858                 if (err) {
1859                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1860                                 h->name, node->dev.id);
1861                         hugetlb_unregister_node(node);
1862                         break;
1863                 }
1864         }
1865 }
1866
1867 /*
1868  * hugetlb init time:  register hstate attributes for all registered node
1869  * devices of nodes that have memory.  All on-line nodes should have
1870  * registered their associated device by this time.
1871  */
1872 static void hugetlb_register_all_nodes(void)
1873 {
1874         int nid;
1875
1876         for_each_node_state(nid, N_MEMORY) {
1877                 struct node *node = node_devices[nid];
1878                 if (node->dev.id == nid)
1879                         hugetlb_register_node(node);
1880         }
1881
1882         /*
1883          * Let the node device driver know we're here so it can
1884          * [un]register hstate attributes on node hotplug.
1885          */
1886         register_hugetlbfs_with_node(hugetlb_register_node,
1887                                      hugetlb_unregister_node);
1888 }
1889 #else   /* !CONFIG_NUMA */
1890
1891 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1892 {
1893         BUG();
1894         if (nidp)
1895                 *nidp = -1;
1896         return NULL;
1897 }
1898
1899 static void hugetlb_unregister_all_nodes(void) { }
1900
1901 static void hugetlb_register_all_nodes(void) { }
1902
1903 #endif
1904
1905 static void __exit hugetlb_exit(void)
1906 {
1907         struct hstate *h;
1908
1909         hugetlb_unregister_all_nodes();
1910
1911         for_each_hstate(h) {
1912                 kobject_put(hstate_kobjs[hstate_index(h)]);
1913         }
1914
1915         kobject_put(hugepages_kobj);
1916 }
1917 module_exit(hugetlb_exit);
1918
1919 static int __init hugetlb_init(void)
1920 {
1921         /* Some platform decide whether they support huge pages at boot
1922          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1923          * there is no such support
1924          */
1925         if (HPAGE_SHIFT == 0)
1926                 return 0;
1927
1928         if (!size_to_hstate(default_hstate_size)) {
1929                 default_hstate_size = HPAGE_SIZE;
1930                 if (!size_to_hstate(default_hstate_size))
1931                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1932         }
1933         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1934         if (default_hstate_max_huge_pages)
1935                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1936
1937         hugetlb_init_hstates();
1938         gather_bootmem_prealloc();
1939         report_hugepages();
1940
1941         hugetlb_sysfs_init();
1942         hugetlb_register_all_nodes();
1943         hugetlb_cgroup_file_init();
1944
1945         return 0;
1946 }
1947 module_init(hugetlb_init);
1948
1949 /* Should be called on processing a hugepagesz=... option */
1950 void __init hugetlb_add_hstate(unsigned order)
1951 {
1952         struct hstate *h;
1953         unsigned long i;
1954
1955         if (size_to_hstate(PAGE_SIZE << order)) {
1956                 pr_warning("hugepagesz= specified twice, ignoring\n");
1957                 return;
1958         }
1959         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1960         BUG_ON(order == 0);
1961         h = &hstates[hugetlb_max_hstate++];
1962         h->order = order;
1963         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1964         h->nr_huge_pages = 0;
1965         h->free_huge_pages = 0;
1966         for (i = 0; i < MAX_NUMNODES; ++i)
1967                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1968         INIT_LIST_HEAD(&h->hugepage_activelist);
1969         h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1970         h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1971         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1972                                         huge_page_size(h)/1024);
1973
1974         parsed_hstate = h;
1975 }
1976
1977 static int __init hugetlb_nrpages_setup(char *s)
1978 {
1979         unsigned long *mhp;
1980         static unsigned long *last_mhp;
1981
1982         /*
1983          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1984          * so this hugepages= parameter goes to the "default hstate".
1985          */
1986         if (!hugetlb_max_hstate)
1987                 mhp = &default_hstate_max_huge_pages;
1988         else
1989                 mhp = &parsed_hstate->max_huge_pages;
1990
1991         if (mhp == last_mhp) {
1992                 pr_warning("hugepages= specified twice without "
1993                            "interleaving hugepagesz=, ignoring\n");
1994                 return 1;
1995         }
1996
1997         if (sscanf(s, "%lu", mhp) <= 0)
1998                 *mhp = 0;
1999
2000         /*
2001          * Global state is always initialized later in hugetlb_init.
2002          * But we need to allocate >= MAX_ORDER hstates here early to still
2003          * use the bootmem allocator.
2004          */
2005         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2006                 hugetlb_hstate_alloc_pages(parsed_hstate);
2007
2008         last_mhp = mhp;
2009
2010         return 1;
2011 }
2012 __setup("hugepages=", hugetlb_nrpages_setup);
2013
2014 static int __init hugetlb_default_setup(char *s)
2015 {
2016         default_hstate_size = memparse(s, &s);
2017         return 1;
2018 }
2019 __setup("default_hugepagesz=", hugetlb_default_setup);
2020
2021 static unsigned int cpuset_mems_nr(unsigned int *array)
2022 {
2023         int node;
2024         unsigned int nr = 0;
2025
2026         for_each_node_mask(node, cpuset_current_mems_allowed)
2027                 nr += array[node];
2028
2029         return nr;
2030 }
2031
2032 #ifdef CONFIG_SYSCTL
2033 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2034                          struct ctl_table *table, int write,
2035                          void __user *buffer, size_t *length, loff_t *ppos)
2036 {
2037         struct hstate *h = &default_hstate;
2038         unsigned long tmp;
2039         int ret;
2040
2041         tmp = h->max_huge_pages;
2042
2043         if (write && h->order >= MAX_ORDER)
2044                 return -EINVAL;
2045
2046         table->data = &tmp;
2047         table->maxlen = sizeof(unsigned long);
2048         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2049         if (ret)
2050                 goto out;
2051
2052         if (write) {
2053                 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2054                                                 GFP_KERNEL | __GFP_NORETRY);
2055                 if (!(obey_mempolicy &&
2056                                init_nodemask_of_mempolicy(nodes_allowed))) {
2057                         NODEMASK_FREE(nodes_allowed);
2058                         nodes_allowed = &node_states[N_MEMORY];
2059                 }
2060                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2061
2062                 if (nodes_allowed != &node_states[N_MEMORY])
2063                         NODEMASK_FREE(nodes_allowed);
2064         }
2065 out:
2066         return ret;
2067 }
2068
2069 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2070                           void __user *buffer, size_t *length, loff_t *ppos)
2071 {
2072
2073         return hugetlb_sysctl_handler_common(false, table, write,
2074                                                         buffer, length, ppos);
2075 }
2076
2077 #ifdef CONFIG_NUMA
2078 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2079                           void __user *buffer, size_t *length, loff_t *ppos)
2080 {
2081         return hugetlb_sysctl_handler_common(true, table, write,
2082                                                         buffer, length, ppos);
2083 }
2084 #endif /* CONFIG_NUMA */
2085
2086 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2087                         void __user *buffer,
2088                         size_t *length, loff_t *ppos)
2089 {
2090         proc_dointvec(table, write, buffer, length, ppos);
2091         if (hugepages_treat_as_movable)
2092                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2093         else
2094                 htlb_alloc_mask = GFP_HIGHUSER;
2095         return 0;
2096 }
2097
2098 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2099                         void __user *buffer,
2100                         size_t *length, loff_t *ppos)
2101 {
2102         struct hstate *h = &default_hstate;
2103         unsigned long tmp;
2104         int ret;
2105
2106         tmp = h->nr_overcommit_huge_pages;
2107
2108         if (write && h->order >= MAX_ORDER)
2109                 return -EINVAL;
2110
2111         table->data = &tmp;
2112         table->maxlen = sizeof(unsigned long);
2113         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2114         if (ret)
2115                 goto out;
2116
2117         if (write) {
2118                 spin_lock(&hugetlb_lock);
2119                 h->nr_overcommit_huge_pages = tmp;
2120                 spin_unlock(&hugetlb_lock);
2121         }
2122 out:
2123         return ret;
2124 }
2125
2126 #endif /* CONFIG_SYSCTL */
2127
2128 void hugetlb_report_meminfo(struct seq_file *m)
2129 {
2130         struct hstate *h = &default_hstate;
2131         seq_printf(m,
2132                         "HugePages_Total:   %5lu\n"
2133                         "HugePages_Free:    %5lu\n"
2134                         "HugePages_Rsvd:    %5lu\n"
2135                         "HugePages_Surp:    %5lu\n"
2136                         "Hugepagesize:   %8lu kB\n",
2137                         h->nr_huge_pages,
2138                         h->free_huge_pages,
2139                         h->resv_huge_pages,
2140                         h->surplus_huge_pages,
2141                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2142 }
2143
2144 int hugetlb_report_node_meminfo(int nid, char *buf)
2145 {
2146         struct hstate *h = &default_hstate;
2147         return sprintf(buf,
2148                 "Node %d HugePages_Total: %5u\n"
2149                 "Node %d HugePages_Free:  %5u\n"
2150                 "Node %d HugePages_Surp:  %5u\n",
2151                 nid, h->nr_huge_pages_node[nid],
2152                 nid, h->free_huge_pages_node[nid],
2153                 nid, h->surplus_huge_pages_node[nid]);
2154 }
2155
2156 void hugetlb_show_meminfo(void)
2157 {
2158         struct hstate *h;
2159         int nid;
2160
2161         for_each_node_state(nid, N_MEMORY)
2162                 for_each_hstate(h)
2163                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2164                                 nid,
2165                                 h->nr_huge_pages_node[nid],
2166                                 h->free_huge_pages_node[nid],
2167                                 h->surplus_huge_pages_node[nid],
2168                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2169 }
2170
2171 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2172 unsigned long hugetlb_total_pages(void)
2173 {
2174         struct hstate *h;
2175         unsigned long nr_total_pages = 0;
2176
2177         for_each_hstate(h)
2178                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2179         return nr_total_pages;
2180 }
2181
2182 static int hugetlb_acct_memory(struct hstate *h, long delta)
2183 {
2184         int ret = -ENOMEM;
2185
2186         spin_lock(&hugetlb_lock);
2187         /*
2188          * When cpuset is configured, it breaks the strict hugetlb page
2189          * reservation as the accounting is done on a global variable. Such
2190          * reservation is completely rubbish in the presence of cpuset because
2191          * the reservation is not checked against page availability for the
2192          * current cpuset. Application can still potentially OOM'ed by kernel
2193          * with lack of free htlb page in cpuset that the task is in.
2194          * Attempt to enforce strict accounting with cpuset is almost
2195          * impossible (or too ugly) because cpuset is too fluid that
2196          * task or memory node can be dynamically moved between cpusets.
2197          *
2198          * The change of semantics for shared hugetlb mapping with cpuset is
2199          * undesirable. However, in order to preserve some of the semantics,
2200          * we fall back to check against current free page availability as
2201          * a best attempt and hopefully to minimize the impact of changing
2202          * semantics that cpuset has.
2203          */
2204         if (delta > 0) {
2205                 if (gather_surplus_pages(h, delta) < 0)
2206                         goto out;
2207
2208                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2209                         return_unused_surplus_pages(h, delta);
2210                         goto out;
2211                 }
2212         }
2213
2214         ret = 0;
2215         if (delta < 0)
2216                 return_unused_surplus_pages(h, (unsigned long) -delta);
2217
2218 out:
2219         spin_unlock(&hugetlb_lock);
2220         return ret;
2221 }
2222
2223 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2224 {
2225         struct resv_map *resv = vma_resv_map(vma);
2226
2227         /*
2228          * This new VMA should share its siblings reservation map if present.
2229          * The VMA will only ever have a valid reservation map pointer where
2230          * it is being copied for another still existing VMA.  As that VMA
2231          * has a reference to the reservation map it cannot disappear until
2232          * after this open call completes.  It is therefore safe to take a
2233          * new reference here without additional locking.
2234          */
2235         if (resv)
2236                 kref_get(&resv->refs);
2237 }
2238
2239 static void resv_map_put(struct vm_area_struct *vma)
2240 {
2241         struct resv_map *resv = vma_resv_map(vma);
2242
2243         if (!resv)
2244                 return;
2245         kref_put(&resv->refs, resv_map_release);
2246 }
2247
2248 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2249 {
2250         struct hstate *h = hstate_vma(vma);
2251         struct resv_map *resv = vma_resv_map(vma);
2252         struct hugepage_subpool *spool = subpool_vma(vma);
2253         unsigned long reserve;
2254         unsigned long start;
2255         unsigned long end;
2256
2257         if (resv) {
2258                 start = vma_hugecache_offset(h, vma, vma->vm_start);
2259                 end = vma_hugecache_offset(h, vma, vma->vm_end);
2260
2261                 reserve = (end - start) -
2262                         region_count(&resv->regions, start, end);
2263
2264                 resv_map_put(vma);
2265
2266                 if (reserve) {
2267                         hugetlb_acct_memory(h, -reserve);
2268                         hugepage_subpool_put_pages(spool, reserve);
2269                 }
2270         }
2271 }
2272
2273 /*
2274  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2275  * handle_mm_fault() to try to instantiate regular-sized pages in the
2276  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2277  * this far.
2278  */
2279 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2280 {
2281         BUG();
2282         return 0;
2283 }
2284
2285 const struct vm_operations_struct hugetlb_vm_ops = {
2286         .fault = hugetlb_vm_op_fault,
2287         .open = hugetlb_vm_op_open,
2288         .close = hugetlb_vm_op_close,
2289 };
2290
2291 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2292                                 int writable)
2293 {
2294         pte_t entry;
2295
2296         if (writable) {
2297                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2298                                          vma->vm_page_prot)));
2299         } else {
2300                 entry = huge_pte_wrprotect(mk_huge_pte(page,
2301                                            vma->vm_page_prot));
2302         }
2303         entry = pte_mkyoung(entry);
2304         entry = pte_mkhuge(entry);
2305         entry = arch_make_huge_pte(entry, vma, page, writable);
2306
2307         return entry;
2308 }
2309
2310 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2311                                    unsigned long address, pte_t *ptep)
2312 {
2313         pte_t entry;
2314
2315         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2316         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2317                 update_mmu_cache(vma, address, ptep);
2318 }
2319
2320 static int is_hugetlb_entry_migration(pte_t pte)
2321 {
2322         swp_entry_t swp;
2323
2324         if (huge_pte_none(pte) || pte_present(pte))
2325                 return 0;
2326         swp = pte_to_swp_entry(pte);
2327         if (non_swap_entry(swp) && is_migration_entry(swp))
2328                 return 1;
2329         else
2330                 return 0;
2331 }
2332
2333 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2334 {
2335         swp_entry_t swp;
2336
2337         if (huge_pte_none(pte) || pte_present(pte))
2338                 return 0;
2339         swp = pte_to_swp_entry(pte);
2340         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2341                 return 1;
2342         else
2343                 return 0;
2344 }
2345
2346 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2347                             struct vm_area_struct *vma)
2348 {
2349         pte_t *src_pte, *dst_pte, entry;
2350         struct page *ptepage;
2351         unsigned long addr;
2352         int cow;
2353         struct hstate *h = hstate_vma(vma);
2354         unsigned long sz = huge_page_size(h);
2355         unsigned long mmun_start;       /* For mmu_notifiers */
2356         unsigned long mmun_end;         /* For mmu_notifiers */
2357         int ret = 0;
2358
2359         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2360
2361         mmun_start = vma->vm_start;
2362         mmun_end = vma->vm_end;
2363         if (cow)
2364                 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2365
2366         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2367                 src_pte = huge_pte_offset(src, addr);
2368                 if (!src_pte)
2369                         continue;
2370                 dst_pte = huge_pte_alloc(dst, addr, sz);
2371                 if (!dst_pte) {
2372                         ret = -ENOMEM;
2373                         break;
2374                 }
2375
2376                 /* If the pagetables are shared don't copy or take references */
2377                 if (dst_pte == src_pte)
2378                         continue;
2379
2380                 spin_lock(&dst->page_table_lock);
2381                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2382                 entry = huge_ptep_get(src_pte);
2383                 if (huge_pte_none(entry)) { /* skip none entry */
2384                         ;
2385                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2386                                     is_hugetlb_entry_hwpoisoned(entry))) {
2387                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
2388
2389                         if (is_write_migration_entry(swp_entry) && cow) {
2390                                 /*
2391                                  * COW mappings require pages in both
2392                                  * parent and child to be set to read.
2393                                  */
2394                                 make_migration_entry_read(&swp_entry);
2395                                 entry = swp_entry_to_pte(swp_entry);
2396                                 set_huge_pte_at(src, addr, src_pte, entry);
2397                         }
2398                         set_huge_pte_at(dst, addr, dst_pte, entry);
2399                 } else {
2400                         if (cow)
2401                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2402                         entry = huge_ptep_get(src_pte);
2403                         ptepage = pte_page(entry);
2404                         get_page(ptepage);
2405                         page_dup_rmap(ptepage);
2406                         set_huge_pte_at(dst, addr, dst_pte, entry);
2407                 }
2408                 spin_unlock(&src->page_table_lock);
2409                 spin_unlock(&dst->page_table_lock);
2410         }
2411
2412         if (cow)
2413                 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2414
2415         return ret;
2416 }
2417
2418 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2419                             unsigned long start, unsigned long end,
2420                             struct page *ref_page)
2421 {
2422         int force_flush = 0;
2423         struct mm_struct *mm = vma->vm_mm;
2424         unsigned long address;
2425         pte_t *ptep;
2426         pte_t pte;
2427         struct page *page;
2428         struct hstate *h = hstate_vma(vma);
2429         unsigned long sz = huge_page_size(h);
2430         const unsigned long mmun_start = start; /* For mmu_notifiers */
2431         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
2432
2433         WARN_ON(!is_vm_hugetlb_page(vma));
2434         BUG_ON(start & ~huge_page_mask(h));
2435         BUG_ON(end & ~huge_page_mask(h));
2436
2437         tlb_start_vma(tlb, vma);
2438         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2439 again:
2440         spin_lock(&mm->page_table_lock);
2441         for (address = start; address < end; address += sz) {
2442                 ptep = huge_pte_offset(mm, address);
2443                 if (!ptep)
2444                         continue;
2445
2446                 if (huge_pmd_unshare(mm, &address, ptep))
2447                         continue;
2448
2449                 pte = huge_ptep_get(ptep);
2450                 if (huge_pte_none(pte))
2451                         continue;
2452
2453                 /*
2454                  * HWPoisoned hugepage is already unmapped and dropped reference
2455                  */
2456                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2457                         huge_pte_clear(mm, address, ptep);
2458                         continue;
2459                 }
2460
2461                 page = pte_page(pte);
2462                 /*
2463                  * If a reference page is supplied, it is because a specific
2464                  * page is being unmapped, not a range. Ensure the page we
2465                  * are about to unmap is the actual page of interest.
2466                  */
2467                 if (ref_page) {
2468                         if (page != ref_page)
2469                                 continue;
2470
2471                         /*
2472                          * Mark the VMA as having unmapped its page so that
2473                          * future faults in this VMA will fail rather than
2474                          * looking like data was lost
2475                          */
2476                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2477                 }
2478
2479                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2480                 tlb_remove_tlb_entry(tlb, ptep, address);
2481                 if (huge_pte_dirty(pte))
2482                         set_page_dirty(page);
2483
2484                 page_remove_rmap(page);
2485                 force_flush = !__tlb_remove_page(tlb, page);
2486                 if (force_flush)
2487                         break;
2488                 /* Bail out after unmapping reference page if supplied */
2489                 if (ref_page)
2490                         break;
2491         }
2492         spin_unlock(&mm->page_table_lock);
2493         /*
2494          * mmu_gather ran out of room to batch pages, we break out of
2495          * the PTE lock to avoid doing the potential expensive TLB invalidate
2496          * and page-free while holding it.
2497          */
2498         if (force_flush) {
2499                 force_flush = 0;
2500                 tlb_flush_mmu(tlb);
2501                 if (address < end && !ref_page)
2502                         goto again;
2503         }
2504         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2505         tlb_end_vma(tlb, vma);
2506 }
2507
2508 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2509                           struct vm_area_struct *vma, unsigned long start,
2510                           unsigned long end, struct page *ref_page)
2511 {
2512         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2513
2514         /*
2515          * Clear this flag so that x86's huge_pmd_share page_table_shareable
2516          * test will fail on a vma being torn down, and not grab a page table
2517          * on its way out.  We're lucky that the flag has such an appropriate
2518          * name, and can in fact be safely cleared here. We could clear it
2519          * before the __unmap_hugepage_range above, but all that's necessary
2520          * is to clear it before releasing the i_mmap_mutex. This works
2521          * because in the context this is called, the VMA is about to be
2522          * destroyed and the i_mmap_mutex is held.
2523          */
2524         vma->vm_flags &= ~VM_MAYSHARE;
2525 }
2526
2527 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2528                           unsigned long end, struct page *ref_page)
2529 {
2530         struct mm_struct *mm;
2531         struct mmu_gather tlb;
2532
2533         mm = vma->vm_mm;
2534
2535         tlb_gather_mmu(&tlb, mm, start, end);
2536         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2537         tlb_finish_mmu(&tlb, start, end);
2538 }
2539
2540 /*
2541  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2542  * mappping it owns the reserve page for. The intention is to unmap the page
2543  * from other VMAs and let the children be SIGKILLed if they are faulting the
2544  * same region.
2545  */
2546 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2547                                 struct page *page, unsigned long address)
2548 {
2549         struct hstate *h = hstate_vma(vma);
2550         struct vm_area_struct *iter_vma;
2551         struct address_space *mapping;
2552         pgoff_t pgoff;
2553
2554         /*
2555          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2556          * from page cache lookup which is in HPAGE_SIZE units.
2557          */
2558         address = address & huge_page_mask(h);
2559         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2560                         vma->vm_pgoff;
2561         mapping = file_inode(vma->vm_file)->i_mapping;
2562
2563         /*
2564          * Take the mapping lock for the duration of the table walk. As
2565          * this mapping should be shared between all the VMAs,
2566          * __unmap_hugepage_range() is called as the lock is already held
2567          */
2568         mutex_lock(&mapping->i_mmap_mutex);
2569         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2570                 /* Do not unmap the current VMA */
2571                 if (iter_vma == vma)
2572                         continue;
2573
2574                 /*
2575                  * Unmap the page from other VMAs without their own reserves.
2576                  * They get marked to be SIGKILLed if they fault in these
2577                  * areas. This is because a future no-page fault on this VMA
2578                  * could insert a zeroed page instead of the data existing
2579                  * from the time of fork. This would look like data corruption
2580                  */
2581                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2582                         unmap_hugepage_range(iter_vma, address,
2583                                              address + huge_page_size(h), page);
2584         }
2585         mutex_unlock(&mapping->i_mmap_mutex);
2586
2587         return 1;
2588 }
2589
2590 /*
2591  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2592  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2593  * cannot race with other handlers or page migration.
2594  * Keep the pte_same checks anyway to make transition from the mutex easier.
2595  */
2596 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2597                         unsigned long address, pte_t *ptep, pte_t pte,
2598                         struct page *pagecache_page)
2599 {
2600         struct hstate *h = hstate_vma(vma);
2601         struct page *old_page, *new_page;
2602         int outside_reserve = 0;
2603         unsigned long mmun_start;       /* For mmu_notifiers */
2604         unsigned long mmun_end;         /* For mmu_notifiers */
2605
2606         old_page = pte_page(pte);
2607
2608 retry_avoidcopy:
2609         /* If no-one else is actually using this page, avoid the copy
2610          * and just make the page writable */
2611         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2612                 page_move_anon_rmap(old_page, vma, address);
2613                 set_huge_ptep_writable(vma, address, ptep);
2614                 return 0;
2615         }
2616
2617         /*
2618          * If the process that created a MAP_PRIVATE mapping is about to
2619          * perform a COW due to a shared page count, attempt to satisfy
2620          * the allocation without using the existing reserves. The pagecache
2621          * page is used to determine if the reserve at this address was
2622          * consumed or not. If reserves were used, a partial faulted mapping
2623          * at the time of fork() could consume its reserves on COW instead
2624          * of the full address range.
2625          */
2626         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2627                         old_page != pagecache_page)
2628                 outside_reserve = 1;
2629
2630         page_cache_get(old_page);
2631
2632         /* Drop page_table_lock as buddy allocator may be called */
2633         spin_unlock(&mm->page_table_lock);
2634         new_page = alloc_huge_page(vma, address, outside_reserve);
2635
2636         if (IS_ERR(new_page)) {
2637                 long err = PTR_ERR(new_page);
2638                 page_cache_release(old_page);
2639
2640                 /*
2641                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2642                  * it is due to references held by a child and an insufficient
2643                  * huge page pool. To guarantee the original mappers
2644                  * reliability, unmap the page from child processes. The child
2645                  * may get SIGKILLed if it later faults.
2646                  */
2647                 if (outside_reserve) {
2648                         BUG_ON(huge_pte_none(pte));
2649                         if (unmap_ref_private(mm, vma, old_page, address)) {
2650                                 BUG_ON(huge_pte_none(pte));
2651                                 spin_lock(&mm->page_table_lock);
2652                                 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2653                                 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2654                                         goto retry_avoidcopy;
2655                                 /*
2656                                  * race occurs while re-acquiring page_table_lock, and
2657                                  * our job is done.
2658                                  */
2659                                 return 0;
2660                         }
2661                         WARN_ON_ONCE(1);
2662                 }
2663
2664                 /* Caller expects lock to be held */
2665                 spin_lock(&mm->page_table_lock);
2666                 if (err == -ENOMEM)
2667                         return VM_FAULT_OOM;
2668                 else
2669                         return VM_FAULT_SIGBUS;
2670         }
2671
2672         /*
2673          * When the original hugepage is shared one, it does not have
2674          * anon_vma prepared.
2675          */
2676         if (unlikely(anon_vma_prepare(vma))) {
2677                 page_cache_release(new_page);
2678                 page_cache_release(old_page);
2679                 /* Caller expects lock to be held */
2680                 spin_lock(&mm->page_table_lock);
2681                 return VM_FAULT_OOM;
2682         }
2683
2684         copy_user_huge_page(new_page, old_page, address, vma,
2685                             pages_per_huge_page(h));
2686         __SetPageUptodate(new_page);
2687
2688         mmun_start = address & huge_page_mask(h);
2689         mmun_end = mmun_start + huge_page_size(h);
2690         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2691         /*
2692          * Retake the page_table_lock to check for racing updates
2693          * before the page tables are altered
2694          */
2695         spin_lock(&mm->page_table_lock);
2696         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2697         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2698                 ClearPagePrivate(new_page);
2699
2700                 /* Break COW */
2701                 huge_ptep_clear_flush(vma, address, ptep);
2702                 set_huge_pte_at(mm, address, ptep,
2703                                 make_huge_pte(vma, new_page, 1));
2704                 page_remove_rmap(old_page);
2705                 hugepage_add_new_anon_rmap(new_page, vma, address);
2706                 /* Make the old page be freed below */
2707                 new_page = old_page;
2708         }
2709         spin_unlock(&mm->page_table_lock);
2710         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2711         page_cache_release(new_page);
2712         page_cache_release(old_page);
2713
2714         /* Caller expects lock to be held */
2715         spin_lock(&mm->page_table_lock);
2716         return 0;
2717 }
2718
2719 /* Return the pagecache page at a given address within a VMA */
2720 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2721                         struct vm_area_struct *vma, unsigned long address)
2722 {
2723         struct address_space *mapping;
2724         pgoff_t idx;
2725
2726         mapping = vma->vm_file->f_mapping;
2727         idx = vma_hugecache_offset(h, vma, address);
2728
2729         return find_lock_page(mapping, idx);
2730 }
2731
2732 /*
2733  * Return whether there is a pagecache page to back given address within VMA.
2734  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2735  */
2736 static bool hugetlbfs_pagecache_present(struct hstate *h,
2737                         struct vm_area_struct *vma, unsigned long address)
2738 {
2739         struct address_space *mapping;
2740         pgoff_t idx;
2741         struct page *page;
2742
2743         mapping = vma->vm_file->f_mapping;
2744         idx = vma_hugecache_offset(h, vma, address);
2745
2746         page = find_get_page(mapping, idx);
2747         if (page)
2748                 put_page(page);
2749         return page != NULL;
2750 }
2751
2752 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2753                         unsigned long address, pte_t *ptep, unsigned int flags)
2754 {
2755         struct hstate *h = hstate_vma(vma);
2756         int ret = VM_FAULT_SIGBUS;
2757         int anon_rmap = 0;
2758         pgoff_t idx;
2759         unsigned long size;
2760         struct page *page;
2761         struct address_space *mapping;
2762         pte_t new_pte;
2763
2764         /*
2765          * Currently, we are forced to kill the process in the event the
2766          * original mapper has unmapped pages from the child due to a failed
2767          * COW. Warn that such a situation has occurred as it may not be obvious
2768          */
2769         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2770                 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2771                            current->pid);
2772                 return ret;
2773         }
2774
2775         mapping = vma->vm_file->f_mapping;
2776         idx = vma_hugecache_offset(h, vma, address);
2777
2778         /*
2779          * Use page lock to guard against racing truncation
2780          * before we get page_table_lock.
2781          */
2782 retry:
2783         page = find_lock_page(mapping, idx);
2784         if (!page) {
2785                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2786                 if (idx >= size)
2787                         goto out;
2788                 page = alloc_huge_page(vma, address, 0);
2789                 if (IS_ERR(page)) {
2790                         ret = PTR_ERR(page);
2791                         if (ret == -ENOMEM)
2792                                 ret = VM_FAULT_OOM;
2793                         else
2794                                 ret = VM_FAULT_SIGBUS;
2795                         goto out;
2796                 }
2797                 clear_huge_page(page, address, pages_per_huge_page(h));
2798                 __SetPageUptodate(page);
2799
2800                 if (vma->vm_flags & VM_MAYSHARE) {
2801                         int err;
2802                         struct inode *inode = mapping->host;
2803
2804                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2805                         if (err) {
2806                                 put_page(page);
2807                                 if (err == -EEXIST)
2808                                         goto retry;
2809                                 goto out;
2810                         }
2811                         ClearPagePrivate(page);
2812
2813                         spin_lock(&inode->i_lock);
2814                         inode->i_blocks += blocks_per_huge_page(h);
2815                         spin_unlock(&inode->i_lock);
2816                 } else {
2817                         lock_page(page);
2818                         if (unlikely(anon_vma_prepare(vma))) {
2819                                 ret = VM_FAULT_OOM;
2820                                 goto backout_unlocked;
2821                         }
2822                         anon_rmap = 1;
2823                 }
2824         } else {
2825                 /*
2826                  * If memory error occurs between mmap() and fault, some process
2827                  * don't have hwpoisoned swap entry for errored virtual address.
2828                  * So we need to block hugepage fault by PG_hwpoison bit check.
2829                  */
2830                 if (unlikely(PageHWPoison(page))) {
2831                         ret = VM_FAULT_HWPOISON |
2832                                 VM_FAULT_SET_HINDEX(hstate_index(h));
2833                         goto backout_unlocked;
2834                 }
2835         }
2836
2837         /*
2838          * If we are going to COW a private mapping later, we examine the
2839          * pending reservations for this page now. This will ensure that
2840          * any allocations necessary to record that reservation occur outside
2841          * the spinlock.
2842          */
2843         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2844                 if (vma_needs_reservation(h, vma, address) < 0) {
2845                         ret = VM_FAULT_OOM;
2846                         goto backout_unlocked;
2847                 }
2848
2849         spin_lock(&mm->page_table_lock);
2850         size = i_size_read(mapping->host) >> huge_page_shift(h);
2851         if (idx >= size)
2852                 goto backout;
2853
2854         ret = 0;
2855         if (!huge_pte_none(huge_ptep_get(ptep)))
2856                 goto backout;
2857
2858         if (anon_rmap) {
2859                 ClearPagePrivate(page);
2860                 hugepage_add_new_anon_rmap(page, vma, address);
2861         }
2862         else
2863                 page_dup_rmap(page);
2864         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2865                                 && (vma->vm_flags & VM_SHARED)));
2866         set_huge_pte_at(mm, address, ptep, new_pte);
2867
2868         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2869                 /* Optimization, do the COW without a second fault */
2870                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2871         }
2872
2873         spin_unlock(&mm->page_table_lock);
2874         unlock_page(page);
2875 out:
2876         return ret;
2877
2878 backout:
2879         spin_unlock(&mm->page_table_lock);
2880 backout_unlocked:
2881         unlock_page(page);
2882         put_page(page);
2883         goto out;
2884 }
2885
2886 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2887                         unsigned long address, unsigned int flags)
2888 {
2889         pte_t *ptep;
2890         pte_t entry;
2891         int ret;
2892         struct page *page = NULL;
2893         struct page *pagecache_page = NULL;
2894         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2895         struct hstate *h = hstate_vma(vma);
2896
2897         address &= huge_page_mask(h);
2898
2899         ptep = huge_pte_offset(mm, address);
2900         if (ptep) {
2901                 entry = huge_ptep_get(ptep);
2902                 if (unlikely(is_hugetlb_entry_migration(entry))) {
2903                         migration_entry_wait_huge(mm, ptep);
2904                         return 0;
2905                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2906                         return VM_FAULT_HWPOISON_LARGE |
2907                                 VM_FAULT_SET_HINDEX(hstate_index(h));
2908         }
2909
2910         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2911         if (!ptep)
2912                 return VM_FAULT_OOM;
2913
2914         /*
2915          * Serialize hugepage allocation and instantiation, so that we don't
2916          * get spurious allocation failures if two CPUs race to instantiate
2917          * the same page in the page cache.
2918          */
2919         mutex_lock(&hugetlb_instantiation_mutex);
2920         entry = huge_ptep_get(ptep);
2921         if (huge_pte_none(entry)) {
2922                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2923                 goto out_mutex;
2924         }
2925
2926         ret = 0;
2927
2928         /*
2929          * If we are going to COW the mapping later, we examine the pending
2930          * reservations for this page now. This will ensure that any
2931          * allocations necessary to record that reservation occur outside the
2932          * spinlock. For private mappings, we also lookup the pagecache
2933          * page now as it is used to determine if a reservation has been
2934          * consumed.
2935          */
2936         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2937                 if (vma_needs_reservation(h, vma, address) < 0) {
2938                         ret = VM_FAULT_OOM;
2939                         goto out_mutex;
2940                 }
2941
2942                 if (!(vma->vm_flags & VM_MAYSHARE))
2943                         pagecache_page = hugetlbfs_pagecache_page(h,
2944                                                                 vma, address);
2945         }
2946
2947         /*
2948          * hugetlb_cow() requires page locks of pte_page(entry) and
2949          * pagecache_page, so here we need take the former one
2950          * when page != pagecache_page or !pagecache_page.
2951          * Note that locking order is always pagecache_page -> page,
2952          * so no worry about deadlock.
2953          */
2954         page = pte_page(entry);
2955         get_page(page);
2956         if (page != pagecache_page)
2957                 lock_page(page);
2958
2959         spin_lock(&mm->page_table_lock);
2960         /* Check for a racing update before calling hugetlb_cow */
2961         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2962                 goto out_page_table_lock;
2963
2964
2965         if (flags & FAULT_FLAG_WRITE) {
2966                 if (!huge_pte_write(entry)) {
2967                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2968                                                         pagecache_page);
2969                         goto out_page_table_lock;
2970                 }
2971                 entry = huge_pte_mkdirty(entry);
2972         }
2973         entry = pte_mkyoung(entry);
2974         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2975                                                 flags & FAULT_FLAG_WRITE))
2976                 update_mmu_cache(vma, address, ptep);
2977
2978 out_page_table_lock:
2979         spin_unlock(&mm->page_table_lock);
2980
2981         if (pagecache_page) {
2982                 unlock_page(pagecache_page);
2983                 put_page(pagecache_page);
2984         }
2985         if (page != pagecache_page)
2986                 unlock_page(page);
2987         put_page(page);
2988
2989 out_mutex:
2990         mutex_unlock(&hugetlb_instantiation_mutex);
2991
2992         return ret;
2993 }
2994
2995 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2996                          struct page **pages, struct vm_area_struct **vmas,
2997                          unsigned long *position, unsigned long *nr_pages,
2998                          long i, unsigned int flags)
2999 {
3000         unsigned long pfn_offset;
3001         unsigned long vaddr = *position;
3002         unsigned long remainder = *nr_pages;
3003         struct hstate *h = hstate_vma(vma);
3004
3005         spin_lock(&mm->page_table_lock);
3006         while (vaddr < vma->vm_end && remainder) {
3007                 pte_t *pte;
3008                 int absent;
3009                 struct page *page;
3010
3011                 /*
3012                  * Some archs (sparc64, sh*) have multiple pte_ts to
3013                  * each hugepage.  We have to make sure we get the
3014                  * first, for the page indexing below to work.
3015                  */
3016                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3017                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3018
3019                 /*
3020                  * When coredumping, it suits get_dump_page if we just return
3021                  * an error where there's an empty slot with no huge pagecache
3022                  * to back it.  This way, we avoid allocating a hugepage, and
3023                  * the sparse dumpfile avoids allocating disk blocks, but its
3024                  * huge holes still show up with zeroes where they need to be.
3025                  */
3026                 if (absent && (flags & FOLL_DUMP) &&
3027                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3028                         remainder = 0;
3029                         break;
3030                 }
3031
3032                 /*
3033                  * We need call hugetlb_fault for both hugepages under migration
3034                  * (in which case hugetlb_fault waits for the migration,) and
3035                  * hwpoisoned hugepages (in which case we need to prevent the
3036                  * caller from accessing to them.) In order to do this, we use
3037                  * here is_swap_pte instead of is_hugetlb_entry_migration and
3038                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3039                  * both cases, and because we can't follow correct pages
3040                  * directly from any kind of swap entries.
3041                  */
3042                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3043                     ((flags & FOLL_WRITE) &&
3044                       !huge_pte_write(huge_ptep_get(pte)))) {
3045                         int ret;
3046
3047                         spin_unlock(&mm->page_table_lock);
3048                         ret = hugetlb_fault(mm, vma, vaddr,
3049                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3050                         spin_lock(&mm->page_table_lock);
3051                         if (!(ret & VM_FAULT_ERROR))
3052                                 continue;
3053
3054                         remainder = 0;
3055                         break;
3056                 }
3057
3058                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3059                 page = pte_page(huge_ptep_get(pte));
3060 same_page:
3061                 if (pages) {
3062                         pages[i] = mem_map_offset(page, pfn_offset);
3063                         get_page(pages[i]);
3064                 }
3065
3066                 if (vmas)
3067                         vmas[i] = vma;
3068
3069                 vaddr += PAGE_SIZE;
3070                 ++pfn_offset;
3071                 --remainder;
3072                 ++i;
3073                 if (vaddr < vma->vm_end && remainder &&
3074                                 pfn_offset < pages_per_huge_page(h)) {
3075                         /*
3076                          * We use pfn_offset to avoid touching the pageframes
3077                          * of this compound page.
3078                          */
3079                         goto same_page;
3080                 }
3081         }
3082         spin_unlock(&mm->page_table_lock);
3083         *nr_pages = remainder;
3084         *position = vaddr;
3085
3086         return i ? i : -EFAULT;
3087 }
3088
3089 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3090                 unsigned long address, unsigned long end, pgprot_t newprot)
3091 {
3092         struct mm_struct *mm = vma->vm_mm;
3093         unsigned long start = address;
3094         pte_t *ptep;
3095         pte_t pte;
3096         struct hstate *h = hstate_vma(vma);
3097         unsigned long pages = 0;
3098
3099         BUG_ON(address >= end);
3100         flush_cache_range(vma, address, end);
3101
3102         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3103         spin_lock(&mm->page_table_lock);
3104         for (; address < end; address += huge_page_size(h)) {
3105                 ptep = huge_pte_offset(mm, address);
3106                 if (!ptep)
3107                         continue;
3108                 if (huge_pmd_unshare(mm, &address, ptep)) {
3109                         pages++;
3110                         continue;
3111                 }
3112                 if (!huge_pte_none(huge_ptep_get(ptep))) {
3113                         pte = huge_ptep_get_and_clear(mm, address, ptep);
3114                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3115                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
3116                         set_huge_pte_at(mm, address, ptep, pte);
3117                         pages++;
3118                 }
3119         }
3120         spin_unlock(&mm->page_table_lock);
3121         /*
3122          * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3123          * may have cleared our pud entry and done put_page on the page table:
3124          * once we release i_mmap_mutex, another task can do the final put_page
3125          * and that page table be reused and filled with junk.
3126          */
3127         flush_tlb_range(vma, start, end);
3128         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3129
3130         return pages << h->order;
3131 }
3132
3133 int hugetlb_reserve_pages(struct inode *inode,
3134                                         long from, long to,
3135                                         struct vm_area_struct *vma,
3136                                         vm_flags_t vm_flags)
3137 {
3138         long ret, chg;
3139         struct hstate *h = hstate_inode(inode);
3140         struct hugepage_subpool *spool = subpool_inode(inode);
3141
3142         /*
3143          * Only apply hugepage reservation if asked. At fault time, an
3144          * attempt will be made for VM_NORESERVE to allocate a page
3145          * without using reserves
3146          */
3147         if (vm_flags & VM_NORESERVE)
3148                 return 0;
3149
3150         /*
3151          * Shared mappings base their reservation on the number of pages that
3152          * are already allocated on behalf of the file. Private mappings need
3153          * to reserve the full area even if read-only as mprotect() may be
3154          * called to make the mapping read-write. Assume !vma is a shm mapping
3155          */
3156         if (!vma || vma->vm_flags & VM_MAYSHARE)
3157                 chg = region_chg(&inode->i_mapping->private_list, from, to);
3158         else {
3159                 struct resv_map *resv_map = resv_map_alloc();
3160                 if (!resv_map)
3161                         return -ENOMEM;
3162
3163                 chg = to - from;
3164
3165                 set_vma_resv_map(vma, resv_map);
3166                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3167         }
3168
3169         if (chg < 0) {
3170                 ret = chg;
3171                 goto out_err;
3172         }
3173
3174         /* There must be enough pages in the subpool for the mapping */
3175         if (hugepage_subpool_get_pages(spool, chg)) {
3176                 ret = -ENOSPC;
3177                 goto out_err;
3178         }
3179
3180         /*
3181          * Check enough hugepages are available for the reservation.
3182          * Hand the pages back to the subpool if there are not
3183          */
3184         ret = hugetlb_acct_memory(h, chg);
3185         if (ret < 0) {
3186                 hugepage_subpool_put_pages(spool, chg);
3187                 goto out_err;
3188         }
3189
3190         /*
3191          * Account for the reservations made. Shared mappings record regions
3192          * that have reservations as they are shared by multiple VMAs.
3193          * When the last VMA disappears, the region map says how much
3194          * the reservation was and the page cache tells how much of
3195          * the reservation was consumed. Private mappings are per-VMA and
3196          * only the consumed reservations are tracked. When the VMA
3197          * disappears, the original reservation is the VMA size and the
3198          * consumed reservations are stored in the map. Hence, nothing
3199          * else has to be done for private mappings here
3200          */
3201         if (!vma || vma->vm_flags & VM_MAYSHARE)
3202                 region_add(&inode->i_mapping->private_list, from, to);
3203         return 0;
3204 out_err:
3205         if (vma)
3206                 resv_map_put(vma);
3207         return ret;
3208 }
3209
3210 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3211 {
3212         struct hstate *h = hstate_inode(inode);
3213         long chg = region_truncate(&inode->i_mapping->private_list, offset);
3214         struct hugepage_subpool *spool = subpool_inode(inode);
3215
3216         spin_lock(&inode->i_lock);
3217         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3218         spin_unlock(&inode->i_lock);
3219
3220         hugepage_subpool_put_pages(spool, (chg - freed));
3221         hugetlb_acct_memory(h, -(chg - freed));
3222 }
3223
3224 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3225 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3226                                 struct vm_area_struct *vma,
3227                                 unsigned long addr, pgoff_t idx)
3228 {
3229         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3230                                 svma->vm_start;
3231         unsigned long sbase = saddr & PUD_MASK;
3232         unsigned long s_end = sbase + PUD_SIZE;
3233
3234         /* Allow segments to share if only one is marked locked */
3235         unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3236         unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3237
3238         /*
3239          * match the virtual addresses, permission and the alignment of the
3240          * page table page.
3241          */
3242         if (pmd_index(addr) != pmd_index(saddr) ||
3243             vm_flags != svm_flags ||
3244             sbase < svma->vm_start || svma->vm_end < s_end)
3245                 return 0;
3246
3247         return saddr;
3248 }
3249
3250 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3251 {
3252         unsigned long base = addr & PUD_MASK;
3253         unsigned long end = base + PUD_SIZE;
3254
3255         /*
3256          * check on proper vm_flags and page table alignment
3257          */
3258         if (vma->vm_flags & VM_MAYSHARE &&
3259             vma->vm_start <= base && end <= vma->vm_end)
3260                 return 1;
3261         return 0;
3262 }
3263
3264 /*
3265  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3266  * and returns the corresponding pte. While this is not necessary for the
3267  * !shared pmd case because we can allocate the pmd later as well, it makes the
3268  * code much cleaner. pmd allocation is essential for the shared case because
3269  * pud has to be populated inside the same i_mmap_mutex section - otherwise
3270  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3271  * bad pmd for sharing.
3272  */
3273 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3274 {
3275         struct vm_area_struct *vma = find_vma(mm, addr);
3276         struct address_space *mapping = vma->vm_file->f_mapping;
3277         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3278                         vma->vm_pgoff;
3279         struct vm_area_struct *svma;
3280         unsigned long saddr;
3281         pte_t *spte = NULL;
3282         pte_t *pte;
3283
3284         if (!vma_shareable(vma, addr))
3285                 return (pte_t *)pmd_alloc(mm, pud, addr);
3286
3287         mutex_lock(&mapping->i_mmap_mutex);
3288         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3289                 if (svma == vma)
3290                         continue;
3291
3292                 saddr = page_table_shareable(svma, vma, addr, idx);
3293                 if (saddr) {
3294                         spte = huge_pte_offset(svma->vm_mm, saddr);
3295                         if (spte) {
3296                                 get_page(virt_to_page(spte));
3297                                 break;
3298                         }
3299                 }
3300         }
3301
3302         if (!spte)
3303                 goto out;
3304
3305         spin_lock(&mm->page_table_lock);
3306         if (pud_none(*pud))
3307                 pud_populate(mm, pud,
3308                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3309         else
3310                 put_page(virt_to_page(spte));
3311         spin_unlock(&mm->page_table_lock);
3312 out:
3313         pte = (pte_t *)pmd_alloc(mm, pud, addr);
3314         mutex_unlock(&mapping->i_mmap_mutex);
3315         return pte;
3316 }
3317
3318 /*
3319  * unmap huge page backed by shared pte.
3320  *
3321  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
3322  * indicated by page_count > 1, unmap is achieved by clearing pud and
3323  * decrementing the ref count. If count == 1, the pte page is not shared.
3324  *
3325  * called with vma->vm_mm->page_table_lock held.
3326  *
3327  * returns: 1 successfully unmapped a shared pte page
3328  *          0 the underlying pte page is not shared, or it is the last user
3329  */
3330 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3331 {
3332         pgd_t *pgd = pgd_offset(mm, *addr);
3333         pud_t *pud = pud_offset(pgd, *addr);
3334
3335         BUG_ON(page_count(virt_to_page(ptep)) == 0);
3336         if (page_count(virt_to_page(ptep)) == 1)
3337                 return 0;
3338
3339         pud_clear(pud);
3340         put_page(virt_to_page(ptep));
3341         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3342         return 1;
3343 }
3344 #define want_pmd_share()        (1)
3345 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3346 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3347 {
3348         return NULL;
3349 }
3350 #define want_pmd_share()        (0)
3351 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3352
3353 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3354 pte_t *huge_pte_alloc(struct mm_struct *mm,
3355                         unsigned long addr, unsigned long sz)
3356 {
3357         pgd_t *pgd;
3358         pud_t *pud;
3359         pte_t *pte = NULL;
3360
3361         pgd = pgd_offset(mm, addr);
3362         pud = pud_alloc(mm, pgd, addr);
3363         if (pud) {
3364                 if (sz == PUD_SIZE) {
3365                         pte = (pte_t *)pud;
3366                 } else {
3367                         BUG_ON(sz != PMD_SIZE);
3368                         if (want_pmd_share() && pud_none(*pud))
3369                                 pte = huge_pmd_share(mm, addr, pud);
3370                         else
3371                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3372                 }
3373         }
3374         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3375
3376         return pte;
3377 }
3378
3379 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3380 {
3381         pgd_t *pgd;
3382         pud_t *pud;
3383         pmd_t *pmd = NULL;
3384
3385         pgd = pgd_offset(mm, addr);
3386         if (pgd_present(*pgd)) {
3387                 pud = pud_offset(pgd, addr);
3388                 if (pud_present(*pud)) {
3389                         if (pud_huge(*pud))
3390                                 return (pte_t *)pud;
3391                         pmd = pmd_offset(pud, addr);
3392                 }
3393         }
3394         return (pte_t *) pmd;
3395 }
3396
3397 struct page *
3398 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3399                 pmd_t *pmd, int write)
3400 {
3401         struct page *page;
3402
3403         page = pte_page(*(pte_t *)pmd);
3404         if (page)
3405                 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3406         return page;
3407 }
3408
3409 struct page *
3410 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3411                 pud_t *pud, int write)
3412 {
3413         struct page *page;
3414
3415         page = pte_page(*(pte_t *)pud);
3416         if (page)
3417                 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3418         return page;
3419 }
3420
3421 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3422
3423 /* Can be overriden by architectures */
3424 __attribute__((weak)) struct page *
3425 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3426                pud_t *pud, int write)
3427 {
3428         BUG();
3429         return NULL;
3430 }
3431
3432 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3433
3434 #ifdef CONFIG_MEMORY_FAILURE
3435
3436 /* Should be called in hugetlb_lock */
3437 static int is_hugepage_on_freelist(struct page *hpage)
3438 {
3439         struct page *page;
3440         struct page *tmp;
3441         struct hstate *h = page_hstate(hpage);
3442         int nid = page_to_nid(hpage);
3443
3444         list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3445                 if (page == hpage)
3446                         return 1;
3447         return 0;
3448 }
3449
3450 /*
3451  * This function is called from memory failure code.
3452  * Assume the caller holds page lock of the head page.
3453  */
3454 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3455 {
3456         struct hstate *h = page_hstate(hpage);
3457         int nid = page_to_nid(hpage);
3458         int ret = -EBUSY;
3459
3460         spin_lock(&hugetlb_lock);
3461         if (is_hugepage_on_freelist(hpage)) {
3462                 /*
3463                  * Hwpoisoned hugepage isn't linked to activelist or freelist,
3464                  * but dangling hpage->lru can trigger list-debug warnings
3465                  * (this happens when we call unpoison_memory() on it),
3466                  * so let it point to itself with list_del_init().
3467                  */
3468                 list_del_init(&hpage->lru);
3469                 set_page_refcounted(hpage);
3470                 h->free_huge_pages--;
3471                 h->free_huge_pages_node[nid]--;
3472                 ret = 0;
3473         }
3474         spin_unlock(&hugetlb_lock);
3475         return ret;
3476 }
3477 #endif