mm: fix ENOSPC returned by handle_mm_fault()
[linux-2.6.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) William Irwin, April 2004
4  */
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24
25 #include <asm/page.h>
26 #include <asm/pgtable.h>
27 #include <asm/io.h>
28
29 #include <linux/hugetlb.h>
30 #include <linux/node.h>
31 #include "internal.h"
32
33 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
36
37 static int max_hstate;
38 unsigned int default_hstate_idx;
39 struct hstate hstates[HUGE_MAX_HSTATE];
40
41 __initdata LIST_HEAD(huge_boot_pages);
42
43 /* for command line parsing */
44 static struct hstate * __initdata parsed_hstate;
45 static unsigned long __initdata default_hstate_max_huge_pages;
46 static unsigned long __initdata default_hstate_size;
47
48 #define for_each_hstate(h) \
49         for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
50
51 /*
52  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53  */
54 static DEFINE_SPINLOCK(hugetlb_lock);
55
56 /*
57  * Region tracking -- allows tracking of reservations and instantiated pages
58  *                    across the pages in a mapping.
59  *
60  * The region data structures are protected by a combination of the mmap_sem
61  * and the hugetlb_instantion_mutex.  To access or modify a region the caller
62  * must either hold the mmap_sem for write, or the mmap_sem for read and
63  * the hugetlb_instantiation mutex:
64  *
65  *      down_write(&mm->mmap_sem);
66  * or
67  *      down_read(&mm->mmap_sem);
68  *      mutex_lock(&hugetlb_instantiation_mutex);
69  */
70 struct file_region {
71         struct list_head link;
72         long from;
73         long to;
74 };
75
76 static long region_add(struct list_head *head, long f, long t)
77 {
78         struct file_region *rg, *nrg, *trg;
79
80         /* Locate the region we are either in or before. */
81         list_for_each_entry(rg, head, link)
82                 if (f <= rg->to)
83                         break;
84
85         /* Round our left edge to the current segment if it encloses us. */
86         if (f > rg->from)
87                 f = rg->from;
88
89         /* Check for and consume any regions we now overlap with. */
90         nrg = rg;
91         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
92                 if (&rg->link == head)
93                         break;
94                 if (rg->from > t)
95                         break;
96
97                 /* If this area reaches higher then extend our area to
98                  * include it completely.  If this is not the first area
99                  * which we intend to reuse, free it. */
100                 if (rg->to > t)
101                         t = rg->to;
102                 if (rg != nrg) {
103                         list_del(&rg->link);
104                         kfree(rg);
105                 }
106         }
107         nrg->from = f;
108         nrg->to = t;
109         return 0;
110 }
111
112 static long region_chg(struct list_head *head, long f, long t)
113 {
114         struct file_region *rg, *nrg;
115         long chg = 0;
116
117         /* Locate the region we are before or in. */
118         list_for_each_entry(rg, head, link)
119                 if (f <= rg->to)
120                         break;
121
122         /* If we are below the current region then a new region is required.
123          * Subtle, allocate a new region at the position but make it zero
124          * size such that we can guarantee to record the reservation. */
125         if (&rg->link == head || t < rg->from) {
126                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
127                 if (!nrg)
128                         return -ENOMEM;
129                 nrg->from = f;
130                 nrg->to   = f;
131                 INIT_LIST_HEAD(&nrg->link);
132                 list_add(&nrg->link, rg->link.prev);
133
134                 return t - f;
135         }
136
137         /* Round our left edge to the current segment if it encloses us. */
138         if (f > rg->from)
139                 f = rg->from;
140         chg = t - f;
141
142         /* Check for and consume any regions we now overlap with. */
143         list_for_each_entry(rg, rg->link.prev, link) {
144                 if (&rg->link == head)
145                         break;
146                 if (rg->from > t)
147                         return chg;
148
149                 /* We overlap with this area, if it extends further than
150                  * us then we must extend ourselves.  Account for its
151                  * existing reservation. */
152                 if (rg->to > t) {
153                         chg += rg->to - t;
154                         t = rg->to;
155                 }
156                 chg -= rg->to - rg->from;
157         }
158         return chg;
159 }
160
161 static long region_truncate(struct list_head *head, long end)
162 {
163         struct file_region *rg, *trg;
164         long chg = 0;
165
166         /* Locate the region we are either in or before. */
167         list_for_each_entry(rg, head, link)
168                 if (end <= rg->to)
169                         break;
170         if (&rg->link == head)
171                 return 0;
172
173         /* If we are in the middle of a region then adjust it. */
174         if (end > rg->from) {
175                 chg = rg->to - end;
176                 rg->to = end;
177                 rg = list_entry(rg->link.next, typeof(*rg), link);
178         }
179
180         /* Drop any remaining regions. */
181         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
182                 if (&rg->link == head)
183                         break;
184                 chg += rg->to - rg->from;
185                 list_del(&rg->link);
186                 kfree(rg);
187         }
188         return chg;
189 }
190
191 static long region_count(struct list_head *head, long f, long t)
192 {
193         struct file_region *rg;
194         long chg = 0;
195
196         /* Locate each segment we overlap with, and count that overlap. */
197         list_for_each_entry(rg, head, link) {
198                 int seg_from;
199                 int seg_to;
200
201                 if (rg->to <= f)
202                         continue;
203                 if (rg->from >= t)
204                         break;
205
206                 seg_from = max(rg->from, f);
207                 seg_to = min(rg->to, t);
208
209                 chg += seg_to - seg_from;
210         }
211
212         return chg;
213 }
214
215 /*
216  * Convert the address within this vma to the page offset within
217  * the mapping, in pagecache page units; huge pages here.
218  */
219 static pgoff_t vma_hugecache_offset(struct hstate *h,
220                         struct vm_area_struct *vma, unsigned long address)
221 {
222         return ((address - vma->vm_start) >> huge_page_shift(h)) +
223                         (vma->vm_pgoff >> huge_page_order(h));
224 }
225
226 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
227                                      unsigned long address)
228 {
229         return vma_hugecache_offset(hstate_vma(vma), vma, address);
230 }
231
232 /*
233  * Return the size of the pages allocated when backing a VMA. In the majority
234  * cases this will be same size as used by the page table entries.
235  */
236 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
237 {
238         struct hstate *hstate;
239
240         if (!is_vm_hugetlb_page(vma))
241                 return PAGE_SIZE;
242
243         hstate = hstate_vma(vma);
244
245         return 1UL << (hstate->order + PAGE_SHIFT);
246 }
247 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
248
249 /*
250  * Return the page size being used by the MMU to back a VMA. In the majority
251  * of cases, the page size used by the kernel matches the MMU size. On
252  * architectures where it differs, an architecture-specific version of this
253  * function is required.
254  */
255 #ifndef vma_mmu_pagesize
256 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
257 {
258         return vma_kernel_pagesize(vma);
259 }
260 #endif
261
262 /*
263  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
264  * bits of the reservation map pointer, which are always clear due to
265  * alignment.
266  */
267 #define HPAGE_RESV_OWNER    (1UL << 0)
268 #define HPAGE_RESV_UNMAPPED (1UL << 1)
269 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
270
271 /*
272  * These helpers are used to track how many pages are reserved for
273  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
274  * is guaranteed to have their future faults succeed.
275  *
276  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
277  * the reserve counters are updated with the hugetlb_lock held. It is safe
278  * to reset the VMA at fork() time as it is not in use yet and there is no
279  * chance of the global counters getting corrupted as a result of the values.
280  *
281  * The private mapping reservation is represented in a subtly different
282  * manner to a shared mapping.  A shared mapping has a region map associated
283  * with the underlying file, this region map represents the backing file
284  * pages which have ever had a reservation assigned which this persists even
285  * after the page is instantiated.  A private mapping has a region map
286  * associated with the original mmap which is attached to all VMAs which
287  * reference it, this region map represents those offsets which have consumed
288  * reservation ie. where pages have been instantiated.
289  */
290 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
291 {
292         return (unsigned long)vma->vm_private_data;
293 }
294
295 static void set_vma_private_data(struct vm_area_struct *vma,
296                                                         unsigned long value)
297 {
298         vma->vm_private_data = (void *)value;
299 }
300
301 struct resv_map {
302         struct kref refs;
303         struct list_head regions;
304 };
305
306 static struct resv_map *resv_map_alloc(void)
307 {
308         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
309         if (!resv_map)
310                 return NULL;
311
312         kref_init(&resv_map->refs);
313         INIT_LIST_HEAD(&resv_map->regions);
314
315         return resv_map;
316 }
317
318 static void resv_map_release(struct kref *ref)
319 {
320         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
321
322         /* Clear out any active regions before we release the map. */
323         region_truncate(&resv_map->regions, 0);
324         kfree(resv_map);
325 }
326
327 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
328 {
329         VM_BUG_ON(!is_vm_hugetlb_page(vma));
330         if (!(vma->vm_flags & VM_MAYSHARE))
331                 return (struct resv_map *)(get_vma_private_data(vma) &
332                                                         ~HPAGE_RESV_MASK);
333         return NULL;
334 }
335
336 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
337 {
338         VM_BUG_ON(!is_vm_hugetlb_page(vma));
339         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
340
341         set_vma_private_data(vma, (get_vma_private_data(vma) &
342                                 HPAGE_RESV_MASK) | (unsigned long)map);
343 }
344
345 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
346 {
347         VM_BUG_ON(!is_vm_hugetlb_page(vma));
348         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
349
350         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
351 }
352
353 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
354 {
355         VM_BUG_ON(!is_vm_hugetlb_page(vma));
356
357         return (get_vma_private_data(vma) & flag) != 0;
358 }
359
360 /* Decrement the reserved pages in the hugepage pool by one */
361 static void decrement_hugepage_resv_vma(struct hstate *h,
362                         struct vm_area_struct *vma)
363 {
364         if (vma->vm_flags & VM_NORESERVE)
365                 return;
366
367         if (vma->vm_flags & VM_MAYSHARE) {
368                 /* Shared mappings always use reserves */
369                 h->resv_huge_pages--;
370         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
371                 /*
372                  * Only the process that called mmap() has reserves for
373                  * private mappings.
374                  */
375                 h->resv_huge_pages--;
376         }
377 }
378
379 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
380 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
381 {
382         VM_BUG_ON(!is_vm_hugetlb_page(vma));
383         if (!(vma->vm_flags & VM_MAYSHARE))
384                 vma->vm_private_data = (void *)0;
385 }
386
387 /* Returns true if the VMA has associated reserve pages */
388 static int vma_has_reserves(struct vm_area_struct *vma)
389 {
390         if (vma->vm_flags & VM_MAYSHARE)
391                 return 1;
392         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
393                 return 1;
394         return 0;
395 }
396
397 static void copy_gigantic_page(struct page *dst, struct page *src)
398 {
399         int i;
400         struct hstate *h = page_hstate(src);
401         struct page *dst_base = dst;
402         struct page *src_base = src;
403
404         for (i = 0; i < pages_per_huge_page(h); ) {
405                 cond_resched();
406                 copy_highpage(dst, src);
407
408                 i++;
409                 dst = mem_map_next(dst, dst_base, i);
410                 src = mem_map_next(src, src_base, i);
411         }
412 }
413
414 void copy_huge_page(struct page *dst, struct page *src)
415 {
416         int i;
417         struct hstate *h = page_hstate(src);
418
419         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
420                 copy_gigantic_page(dst, src);
421                 return;
422         }
423
424         might_sleep();
425         for (i = 0; i < pages_per_huge_page(h); i++) {
426                 cond_resched();
427                 copy_highpage(dst + i, src + i);
428         }
429 }
430
431 static void enqueue_huge_page(struct hstate *h, struct page *page)
432 {
433         int nid = page_to_nid(page);
434         list_add(&page->lru, &h->hugepage_freelists[nid]);
435         h->free_huge_pages++;
436         h->free_huge_pages_node[nid]++;
437 }
438
439 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
440 {
441         struct page *page;
442
443         if (list_empty(&h->hugepage_freelists[nid]))
444                 return NULL;
445         page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
446         list_del(&page->lru);
447         set_page_refcounted(page);
448         h->free_huge_pages--;
449         h->free_huge_pages_node[nid]--;
450         return page;
451 }
452
453 static struct page *dequeue_huge_page_vma(struct hstate *h,
454                                 struct vm_area_struct *vma,
455                                 unsigned long address, int avoid_reserve)
456 {
457         struct page *page = NULL;
458         struct mempolicy *mpol;
459         nodemask_t *nodemask;
460         struct zonelist *zonelist;
461         struct zone *zone;
462         struct zoneref *z;
463
464         get_mems_allowed();
465         zonelist = huge_zonelist(vma, address,
466                                         htlb_alloc_mask, &mpol, &nodemask);
467         /*
468          * A child process with MAP_PRIVATE mappings created by their parent
469          * have no page reserves. This check ensures that reservations are
470          * not "stolen". The child may still get SIGKILLed
471          */
472         if (!vma_has_reserves(vma) &&
473                         h->free_huge_pages - h->resv_huge_pages == 0)
474                 goto err;
475
476         /* If reserves cannot be used, ensure enough pages are in the pool */
477         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
478                 goto err;
479
480         for_each_zone_zonelist_nodemask(zone, z, zonelist,
481                                                 MAX_NR_ZONES - 1, nodemask) {
482                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
483                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
484                         if (page) {
485                                 if (!avoid_reserve)
486                                         decrement_hugepage_resv_vma(h, vma);
487                                 break;
488                         }
489                 }
490         }
491 err:
492         mpol_cond_put(mpol);
493         put_mems_allowed();
494         return page;
495 }
496
497 static void update_and_free_page(struct hstate *h, struct page *page)
498 {
499         int i;
500
501         VM_BUG_ON(h->order >= MAX_ORDER);
502
503         h->nr_huge_pages--;
504         h->nr_huge_pages_node[page_to_nid(page)]--;
505         for (i = 0; i < pages_per_huge_page(h); i++) {
506                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
507                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
508                                 1 << PG_private | 1<< PG_writeback);
509         }
510         set_compound_page_dtor(page, NULL);
511         set_page_refcounted(page);
512         arch_release_hugepage(page);
513         __free_pages(page, huge_page_order(h));
514 }
515
516 struct hstate *size_to_hstate(unsigned long size)
517 {
518         struct hstate *h;
519
520         for_each_hstate(h) {
521                 if (huge_page_size(h) == size)
522                         return h;
523         }
524         return NULL;
525 }
526
527 static void free_huge_page(struct page *page)
528 {
529         /*
530          * Can't pass hstate in here because it is called from the
531          * compound page destructor.
532          */
533         struct hstate *h = page_hstate(page);
534         int nid = page_to_nid(page);
535         struct address_space *mapping;
536
537         mapping = (struct address_space *) page_private(page);
538         set_page_private(page, 0);
539         page->mapping = NULL;
540         BUG_ON(page_count(page));
541         BUG_ON(page_mapcount(page));
542         INIT_LIST_HEAD(&page->lru);
543
544         spin_lock(&hugetlb_lock);
545         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
546                 update_and_free_page(h, page);
547                 h->surplus_huge_pages--;
548                 h->surplus_huge_pages_node[nid]--;
549         } else {
550                 enqueue_huge_page(h, page);
551         }
552         spin_unlock(&hugetlb_lock);
553         if (mapping)
554                 hugetlb_put_quota(mapping, 1);
555 }
556
557 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
558 {
559         set_compound_page_dtor(page, free_huge_page);
560         spin_lock(&hugetlb_lock);
561         h->nr_huge_pages++;
562         h->nr_huge_pages_node[nid]++;
563         spin_unlock(&hugetlb_lock);
564         put_page(page); /* free it into the hugepage allocator */
565 }
566
567 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
568 {
569         int i;
570         int nr_pages = 1 << order;
571         struct page *p = page + 1;
572
573         /* we rely on prep_new_huge_page to set the destructor */
574         set_compound_order(page, order);
575         __SetPageHead(page);
576         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
577                 __SetPageTail(p);
578                 p->first_page = page;
579         }
580 }
581
582 int PageHuge(struct page *page)
583 {
584         compound_page_dtor *dtor;
585
586         if (!PageCompound(page))
587                 return 0;
588
589         page = compound_head(page);
590         dtor = get_compound_page_dtor(page);
591
592         return dtor == free_huge_page;
593 }
594
595 EXPORT_SYMBOL_GPL(PageHuge);
596
597 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
598 {
599         struct page *page;
600
601         if (h->order >= MAX_ORDER)
602                 return NULL;
603
604         page = alloc_pages_exact_node(nid,
605                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
606                                                 __GFP_REPEAT|__GFP_NOWARN,
607                 huge_page_order(h));
608         if (page) {
609                 if (arch_prepare_hugepage(page)) {
610                         __free_pages(page, huge_page_order(h));
611                         return NULL;
612                 }
613                 prep_new_huge_page(h, page, nid);
614         }
615
616         return page;
617 }
618
619 /*
620  * common helper functions for hstate_next_node_to_{alloc|free}.
621  * We may have allocated or freed a huge page based on a different
622  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
623  * be outside of *nodes_allowed.  Ensure that we use an allowed
624  * node for alloc or free.
625  */
626 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
627 {
628         nid = next_node(nid, *nodes_allowed);
629         if (nid == MAX_NUMNODES)
630                 nid = first_node(*nodes_allowed);
631         VM_BUG_ON(nid >= MAX_NUMNODES);
632
633         return nid;
634 }
635
636 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
637 {
638         if (!node_isset(nid, *nodes_allowed))
639                 nid = next_node_allowed(nid, nodes_allowed);
640         return nid;
641 }
642
643 /*
644  * returns the previously saved node ["this node"] from which to
645  * allocate a persistent huge page for the pool and advance the
646  * next node from which to allocate, handling wrap at end of node
647  * mask.
648  */
649 static int hstate_next_node_to_alloc(struct hstate *h,
650                                         nodemask_t *nodes_allowed)
651 {
652         int nid;
653
654         VM_BUG_ON(!nodes_allowed);
655
656         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
657         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
658
659         return nid;
660 }
661
662 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
663 {
664         struct page *page;
665         int start_nid;
666         int next_nid;
667         int ret = 0;
668
669         start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
670         next_nid = start_nid;
671
672         do {
673                 page = alloc_fresh_huge_page_node(h, next_nid);
674                 if (page) {
675                         ret = 1;
676                         break;
677                 }
678                 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
679         } while (next_nid != start_nid);
680
681         if (ret)
682                 count_vm_event(HTLB_BUDDY_PGALLOC);
683         else
684                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
685
686         return ret;
687 }
688
689 /*
690  * helper for free_pool_huge_page() - return the previously saved
691  * node ["this node"] from which to free a huge page.  Advance the
692  * next node id whether or not we find a free huge page to free so
693  * that the next attempt to free addresses the next node.
694  */
695 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
696 {
697         int nid;
698
699         VM_BUG_ON(!nodes_allowed);
700
701         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
702         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
703
704         return nid;
705 }
706
707 /*
708  * Free huge page from pool from next node to free.
709  * Attempt to keep persistent huge pages more or less
710  * balanced over allowed nodes.
711  * Called with hugetlb_lock locked.
712  */
713 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
714                                                          bool acct_surplus)
715 {
716         int start_nid;
717         int next_nid;
718         int ret = 0;
719
720         start_nid = hstate_next_node_to_free(h, nodes_allowed);
721         next_nid = start_nid;
722
723         do {
724                 /*
725                  * If we're returning unused surplus pages, only examine
726                  * nodes with surplus pages.
727                  */
728                 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
729                     !list_empty(&h->hugepage_freelists[next_nid])) {
730                         struct page *page =
731                                 list_entry(h->hugepage_freelists[next_nid].next,
732                                           struct page, lru);
733                         list_del(&page->lru);
734                         h->free_huge_pages--;
735                         h->free_huge_pages_node[next_nid]--;
736                         if (acct_surplus) {
737                                 h->surplus_huge_pages--;
738                                 h->surplus_huge_pages_node[next_nid]--;
739                         }
740                         update_and_free_page(h, page);
741                         ret = 1;
742                         break;
743                 }
744                 next_nid = hstate_next_node_to_free(h, nodes_allowed);
745         } while (next_nid != start_nid);
746
747         return ret;
748 }
749
750 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
751 {
752         struct page *page;
753         unsigned int r_nid;
754
755         if (h->order >= MAX_ORDER)
756                 return NULL;
757
758         /*
759          * Assume we will successfully allocate the surplus page to
760          * prevent racing processes from causing the surplus to exceed
761          * overcommit
762          *
763          * This however introduces a different race, where a process B
764          * tries to grow the static hugepage pool while alloc_pages() is
765          * called by process A. B will only examine the per-node
766          * counters in determining if surplus huge pages can be
767          * converted to normal huge pages in adjust_pool_surplus(). A
768          * won't be able to increment the per-node counter, until the
769          * lock is dropped by B, but B doesn't drop hugetlb_lock until
770          * no more huge pages can be converted from surplus to normal
771          * state (and doesn't try to convert again). Thus, we have a
772          * case where a surplus huge page exists, the pool is grown, and
773          * the surplus huge page still exists after, even though it
774          * should just have been converted to a normal huge page. This
775          * does not leak memory, though, as the hugepage will be freed
776          * once it is out of use. It also does not allow the counters to
777          * go out of whack in adjust_pool_surplus() as we don't modify
778          * the node values until we've gotten the hugepage and only the
779          * per-node value is checked there.
780          */
781         spin_lock(&hugetlb_lock);
782         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
783                 spin_unlock(&hugetlb_lock);
784                 return NULL;
785         } else {
786                 h->nr_huge_pages++;
787                 h->surplus_huge_pages++;
788         }
789         spin_unlock(&hugetlb_lock);
790
791         if (nid == NUMA_NO_NODE)
792                 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
793                                    __GFP_REPEAT|__GFP_NOWARN,
794                                    huge_page_order(h));
795         else
796                 page = alloc_pages_exact_node(nid,
797                         htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
798                         __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
799
800         if (page && arch_prepare_hugepage(page)) {
801                 __free_pages(page, huge_page_order(h));
802                 return NULL;
803         }
804
805         spin_lock(&hugetlb_lock);
806         if (page) {
807                 r_nid = page_to_nid(page);
808                 set_compound_page_dtor(page, free_huge_page);
809                 /*
810                  * We incremented the global counters already
811                  */
812                 h->nr_huge_pages_node[r_nid]++;
813                 h->surplus_huge_pages_node[r_nid]++;
814                 __count_vm_event(HTLB_BUDDY_PGALLOC);
815         } else {
816                 h->nr_huge_pages--;
817                 h->surplus_huge_pages--;
818                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
819         }
820         spin_unlock(&hugetlb_lock);
821
822         return page;
823 }
824
825 /*
826  * This allocation function is useful in the context where vma is irrelevant.
827  * E.g. soft-offlining uses this function because it only cares physical
828  * address of error page.
829  */
830 struct page *alloc_huge_page_node(struct hstate *h, int nid)
831 {
832         struct page *page;
833
834         spin_lock(&hugetlb_lock);
835         page = dequeue_huge_page_node(h, nid);
836         spin_unlock(&hugetlb_lock);
837
838         if (!page)
839                 page = alloc_buddy_huge_page(h, nid);
840
841         return page;
842 }
843
844 /*
845  * Increase the hugetlb pool such that it can accommodate a reservation
846  * of size 'delta'.
847  */
848 static int gather_surplus_pages(struct hstate *h, int delta)
849 {
850         struct list_head surplus_list;
851         struct page *page, *tmp;
852         int ret, i;
853         int needed, allocated;
854
855         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
856         if (needed <= 0) {
857                 h->resv_huge_pages += delta;
858                 return 0;
859         }
860
861         allocated = 0;
862         INIT_LIST_HEAD(&surplus_list);
863
864         ret = -ENOMEM;
865 retry:
866         spin_unlock(&hugetlb_lock);
867         for (i = 0; i < needed; i++) {
868                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
869                 if (!page)
870                         /*
871                          * We were not able to allocate enough pages to
872                          * satisfy the entire reservation so we free what
873                          * we've allocated so far.
874                          */
875                         goto free;
876
877                 list_add(&page->lru, &surplus_list);
878         }
879         allocated += needed;
880
881         /*
882          * After retaking hugetlb_lock, we need to recalculate 'needed'
883          * because either resv_huge_pages or free_huge_pages may have changed.
884          */
885         spin_lock(&hugetlb_lock);
886         needed = (h->resv_huge_pages + delta) -
887                         (h->free_huge_pages + allocated);
888         if (needed > 0)
889                 goto retry;
890
891         /*
892          * The surplus_list now contains _at_least_ the number of extra pages
893          * needed to accommodate the reservation.  Add the appropriate number
894          * of pages to the hugetlb pool and free the extras back to the buddy
895          * allocator.  Commit the entire reservation here to prevent another
896          * process from stealing the pages as they are added to the pool but
897          * before they are reserved.
898          */
899         needed += allocated;
900         h->resv_huge_pages += delta;
901         ret = 0;
902
903         spin_unlock(&hugetlb_lock);
904         /* Free the needed pages to the hugetlb pool */
905         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
906                 if ((--needed) < 0)
907                         break;
908                 list_del(&page->lru);
909                 /*
910                  * This page is now managed by the hugetlb allocator and has
911                  * no users -- drop the buddy allocator's reference.
912                  */
913                 put_page_testzero(page);
914                 VM_BUG_ON(page_count(page));
915                 enqueue_huge_page(h, page);
916         }
917
918         /* Free unnecessary surplus pages to the buddy allocator */
919 free:
920         if (!list_empty(&surplus_list)) {
921                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
922                         list_del(&page->lru);
923                         put_page(page);
924                 }
925         }
926         spin_lock(&hugetlb_lock);
927
928         return ret;
929 }
930
931 /*
932  * When releasing a hugetlb pool reservation, any surplus pages that were
933  * allocated to satisfy the reservation must be explicitly freed if they were
934  * never used.
935  * Called with hugetlb_lock held.
936  */
937 static void return_unused_surplus_pages(struct hstate *h,
938                                         unsigned long unused_resv_pages)
939 {
940         unsigned long nr_pages;
941
942         /* Uncommit the reservation */
943         h->resv_huge_pages -= unused_resv_pages;
944
945         /* Cannot return gigantic pages currently */
946         if (h->order >= MAX_ORDER)
947                 return;
948
949         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
950
951         /*
952          * We want to release as many surplus pages as possible, spread
953          * evenly across all nodes with memory. Iterate across these nodes
954          * until we can no longer free unreserved surplus pages. This occurs
955          * when the nodes with surplus pages have no free pages.
956          * free_pool_huge_page() will balance the the freed pages across the
957          * on-line nodes with memory and will handle the hstate accounting.
958          */
959         while (nr_pages--) {
960                 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
961                         break;
962         }
963 }
964
965 /*
966  * Determine if the huge page at addr within the vma has an associated
967  * reservation.  Where it does not we will need to logically increase
968  * reservation and actually increase quota before an allocation can occur.
969  * Where any new reservation would be required the reservation change is
970  * prepared, but not committed.  Once the page has been quota'd allocated
971  * an instantiated the change should be committed via vma_commit_reservation.
972  * No action is required on failure.
973  */
974 static long vma_needs_reservation(struct hstate *h,
975                         struct vm_area_struct *vma, unsigned long addr)
976 {
977         struct address_space *mapping = vma->vm_file->f_mapping;
978         struct inode *inode = mapping->host;
979
980         if (vma->vm_flags & VM_MAYSHARE) {
981                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
982                 return region_chg(&inode->i_mapping->private_list,
983                                                         idx, idx + 1);
984
985         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
986                 return 1;
987
988         } else  {
989                 long err;
990                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
991                 struct resv_map *reservations = vma_resv_map(vma);
992
993                 err = region_chg(&reservations->regions, idx, idx + 1);
994                 if (err < 0)
995                         return err;
996                 return 0;
997         }
998 }
999 static void vma_commit_reservation(struct hstate *h,
1000                         struct vm_area_struct *vma, unsigned long addr)
1001 {
1002         struct address_space *mapping = vma->vm_file->f_mapping;
1003         struct inode *inode = mapping->host;
1004
1005         if (vma->vm_flags & VM_MAYSHARE) {
1006                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1007                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1008
1009         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1010                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1011                 struct resv_map *reservations = vma_resv_map(vma);
1012
1013                 /* Mark this page used in the map. */
1014                 region_add(&reservations->regions, idx, idx + 1);
1015         }
1016 }
1017
1018 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1019                                     unsigned long addr, int avoid_reserve)
1020 {
1021         struct hstate *h = hstate_vma(vma);
1022         struct page *page;
1023         struct address_space *mapping = vma->vm_file->f_mapping;
1024         struct inode *inode = mapping->host;
1025         long chg;
1026
1027         /*
1028          * Processes that did not create the mapping will have no reserves and
1029          * will not have accounted against quota. Check that the quota can be
1030          * made before satisfying the allocation
1031          * MAP_NORESERVE mappings may also need pages and quota allocated
1032          * if no reserve mapping overlaps.
1033          */
1034         chg = vma_needs_reservation(h, vma, addr);
1035         if (chg < 0)
1036                 return ERR_PTR(-VM_FAULT_OOM);
1037         if (chg)
1038                 if (hugetlb_get_quota(inode->i_mapping, chg))
1039                         return ERR_PTR(-VM_FAULT_SIGBUS);
1040
1041         spin_lock(&hugetlb_lock);
1042         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1043         spin_unlock(&hugetlb_lock);
1044
1045         if (!page) {
1046                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1047                 if (!page) {
1048                         hugetlb_put_quota(inode->i_mapping, chg);
1049                         return ERR_PTR(-VM_FAULT_SIGBUS);
1050                 }
1051         }
1052
1053         set_page_private(page, (unsigned long) mapping);
1054
1055         vma_commit_reservation(h, vma, addr);
1056
1057         return page;
1058 }
1059
1060 int __weak alloc_bootmem_huge_page(struct hstate *h)
1061 {
1062         struct huge_bootmem_page *m;
1063         int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1064
1065         while (nr_nodes) {
1066                 void *addr;
1067
1068                 addr = __alloc_bootmem_node_nopanic(
1069                                 NODE_DATA(hstate_next_node_to_alloc(h,
1070                                                 &node_states[N_HIGH_MEMORY])),
1071                                 huge_page_size(h), huge_page_size(h), 0);
1072
1073                 if (addr) {
1074                         /*
1075                          * Use the beginning of the huge page to store the
1076                          * huge_bootmem_page struct (until gather_bootmem
1077                          * puts them into the mem_map).
1078                          */
1079                         m = addr;
1080                         goto found;
1081                 }
1082                 nr_nodes--;
1083         }
1084         return 0;
1085
1086 found:
1087         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1088         /* Put them into a private list first because mem_map is not up yet */
1089         list_add(&m->list, &huge_boot_pages);
1090         m->hstate = h;
1091         return 1;
1092 }
1093
1094 static void prep_compound_huge_page(struct page *page, int order)
1095 {
1096         if (unlikely(order > (MAX_ORDER - 1)))
1097                 prep_compound_gigantic_page(page, order);
1098         else
1099                 prep_compound_page(page, order);
1100 }
1101
1102 /* Put bootmem huge pages into the standard lists after mem_map is up */
1103 static void __init gather_bootmem_prealloc(void)
1104 {
1105         struct huge_bootmem_page *m;
1106
1107         list_for_each_entry(m, &huge_boot_pages, list) {
1108                 struct page *page = virt_to_page(m);
1109                 struct hstate *h = m->hstate;
1110                 __ClearPageReserved(page);
1111                 WARN_ON(page_count(page) != 1);
1112                 prep_compound_huge_page(page, h->order);
1113                 prep_new_huge_page(h, page, page_to_nid(page));
1114         }
1115 }
1116
1117 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1118 {
1119         unsigned long i;
1120
1121         for (i = 0; i < h->max_huge_pages; ++i) {
1122                 if (h->order >= MAX_ORDER) {
1123                         if (!alloc_bootmem_huge_page(h))
1124                                 break;
1125                 } else if (!alloc_fresh_huge_page(h,
1126                                          &node_states[N_HIGH_MEMORY]))
1127                         break;
1128         }
1129         h->max_huge_pages = i;
1130 }
1131
1132 static void __init hugetlb_init_hstates(void)
1133 {
1134         struct hstate *h;
1135
1136         for_each_hstate(h) {
1137                 /* oversize hugepages were init'ed in early boot */
1138                 if (h->order < MAX_ORDER)
1139                         hugetlb_hstate_alloc_pages(h);
1140         }
1141 }
1142
1143 static char * __init memfmt(char *buf, unsigned long n)
1144 {
1145         if (n >= (1UL << 30))
1146                 sprintf(buf, "%lu GB", n >> 30);
1147         else if (n >= (1UL << 20))
1148                 sprintf(buf, "%lu MB", n >> 20);
1149         else
1150                 sprintf(buf, "%lu KB", n >> 10);
1151         return buf;
1152 }
1153
1154 static void __init report_hugepages(void)
1155 {
1156         struct hstate *h;
1157
1158         for_each_hstate(h) {
1159                 char buf[32];
1160                 printk(KERN_INFO "HugeTLB registered %s page size, "
1161                                  "pre-allocated %ld pages\n",
1162                         memfmt(buf, huge_page_size(h)),
1163                         h->free_huge_pages);
1164         }
1165 }
1166
1167 #ifdef CONFIG_HIGHMEM
1168 static void try_to_free_low(struct hstate *h, unsigned long count,
1169                                                 nodemask_t *nodes_allowed)
1170 {
1171         int i;
1172
1173         if (h->order >= MAX_ORDER)
1174                 return;
1175
1176         for_each_node_mask(i, *nodes_allowed) {
1177                 struct page *page, *next;
1178                 struct list_head *freel = &h->hugepage_freelists[i];
1179                 list_for_each_entry_safe(page, next, freel, lru) {
1180                         if (count >= h->nr_huge_pages)
1181                                 return;
1182                         if (PageHighMem(page))
1183                                 continue;
1184                         list_del(&page->lru);
1185                         update_and_free_page(h, page);
1186                         h->free_huge_pages--;
1187                         h->free_huge_pages_node[page_to_nid(page)]--;
1188                 }
1189         }
1190 }
1191 #else
1192 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1193                                                 nodemask_t *nodes_allowed)
1194 {
1195 }
1196 #endif
1197
1198 /*
1199  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1200  * balanced by operating on them in a round-robin fashion.
1201  * Returns 1 if an adjustment was made.
1202  */
1203 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1204                                 int delta)
1205 {
1206         int start_nid, next_nid;
1207         int ret = 0;
1208
1209         VM_BUG_ON(delta != -1 && delta != 1);
1210
1211         if (delta < 0)
1212                 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1213         else
1214                 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1215         next_nid = start_nid;
1216
1217         do {
1218                 int nid = next_nid;
1219                 if (delta < 0)  {
1220                         /*
1221                          * To shrink on this node, there must be a surplus page
1222                          */
1223                         if (!h->surplus_huge_pages_node[nid]) {
1224                                 next_nid = hstate_next_node_to_alloc(h,
1225                                                                 nodes_allowed);
1226                                 continue;
1227                         }
1228                 }
1229                 if (delta > 0) {
1230                         /*
1231                          * Surplus cannot exceed the total number of pages
1232                          */
1233                         if (h->surplus_huge_pages_node[nid] >=
1234                                                 h->nr_huge_pages_node[nid]) {
1235                                 next_nid = hstate_next_node_to_free(h,
1236                                                                 nodes_allowed);
1237                                 continue;
1238                         }
1239                 }
1240
1241                 h->surplus_huge_pages += delta;
1242                 h->surplus_huge_pages_node[nid] += delta;
1243                 ret = 1;
1244                 break;
1245         } while (next_nid != start_nid);
1246
1247         return ret;
1248 }
1249
1250 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1251 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1252                                                 nodemask_t *nodes_allowed)
1253 {
1254         unsigned long min_count, ret;
1255
1256         if (h->order >= MAX_ORDER)
1257                 return h->max_huge_pages;
1258
1259         /*
1260          * Increase the pool size
1261          * First take pages out of surplus state.  Then make up the
1262          * remaining difference by allocating fresh huge pages.
1263          *
1264          * We might race with alloc_buddy_huge_page() here and be unable
1265          * to convert a surplus huge page to a normal huge page. That is
1266          * not critical, though, it just means the overall size of the
1267          * pool might be one hugepage larger than it needs to be, but
1268          * within all the constraints specified by the sysctls.
1269          */
1270         spin_lock(&hugetlb_lock);
1271         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1272                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1273                         break;
1274         }
1275
1276         while (count > persistent_huge_pages(h)) {
1277                 /*
1278                  * If this allocation races such that we no longer need the
1279                  * page, free_huge_page will handle it by freeing the page
1280                  * and reducing the surplus.
1281                  */
1282                 spin_unlock(&hugetlb_lock);
1283                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1284                 spin_lock(&hugetlb_lock);
1285                 if (!ret)
1286                         goto out;
1287
1288                 /* Bail for signals. Probably ctrl-c from user */
1289                 if (signal_pending(current))
1290                         goto out;
1291         }
1292
1293         /*
1294          * Decrease the pool size
1295          * First return free pages to the buddy allocator (being careful
1296          * to keep enough around to satisfy reservations).  Then place
1297          * pages into surplus state as needed so the pool will shrink
1298          * to the desired size as pages become free.
1299          *
1300          * By placing pages into the surplus state independent of the
1301          * overcommit value, we are allowing the surplus pool size to
1302          * exceed overcommit. There are few sane options here. Since
1303          * alloc_buddy_huge_page() is checking the global counter,
1304          * though, we'll note that we're not allowed to exceed surplus
1305          * and won't grow the pool anywhere else. Not until one of the
1306          * sysctls are changed, or the surplus pages go out of use.
1307          */
1308         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1309         min_count = max(count, min_count);
1310         try_to_free_low(h, min_count, nodes_allowed);
1311         while (min_count < persistent_huge_pages(h)) {
1312                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1313                         break;
1314         }
1315         while (count < persistent_huge_pages(h)) {
1316                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1317                         break;
1318         }
1319 out:
1320         ret = persistent_huge_pages(h);
1321         spin_unlock(&hugetlb_lock);
1322         return ret;
1323 }
1324
1325 #define HSTATE_ATTR_RO(_name) \
1326         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1327
1328 #define HSTATE_ATTR(_name) \
1329         static struct kobj_attribute _name##_attr = \
1330                 __ATTR(_name, 0644, _name##_show, _name##_store)
1331
1332 static struct kobject *hugepages_kobj;
1333 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1334
1335 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1336
1337 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1338 {
1339         int i;
1340
1341         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1342                 if (hstate_kobjs[i] == kobj) {
1343                         if (nidp)
1344                                 *nidp = NUMA_NO_NODE;
1345                         return &hstates[i];
1346                 }
1347
1348         return kobj_to_node_hstate(kobj, nidp);
1349 }
1350
1351 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1352                                         struct kobj_attribute *attr, char *buf)
1353 {
1354         struct hstate *h;
1355         unsigned long nr_huge_pages;
1356         int nid;
1357
1358         h = kobj_to_hstate(kobj, &nid);
1359         if (nid == NUMA_NO_NODE)
1360                 nr_huge_pages = h->nr_huge_pages;
1361         else
1362                 nr_huge_pages = h->nr_huge_pages_node[nid];
1363
1364         return sprintf(buf, "%lu\n", nr_huge_pages);
1365 }
1366
1367 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1368                         struct kobject *kobj, struct kobj_attribute *attr,
1369                         const char *buf, size_t len)
1370 {
1371         int err;
1372         int nid;
1373         unsigned long count;
1374         struct hstate *h;
1375         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1376
1377         err = strict_strtoul(buf, 10, &count);
1378         if (err)
1379                 goto out;
1380
1381         h = kobj_to_hstate(kobj, &nid);
1382         if (h->order >= MAX_ORDER) {
1383                 err = -EINVAL;
1384                 goto out;
1385         }
1386
1387         if (nid == NUMA_NO_NODE) {
1388                 /*
1389                  * global hstate attribute
1390                  */
1391                 if (!(obey_mempolicy &&
1392                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1393                         NODEMASK_FREE(nodes_allowed);
1394                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1395                 }
1396         } else if (nodes_allowed) {
1397                 /*
1398                  * per node hstate attribute: adjust count to global,
1399                  * but restrict alloc/free to the specified node.
1400                  */
1401                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1402                 init_nodemask_of_node(nodes_allowed, nid);
1403         } else
1404                 nodes_allowed = &node_states[N_HIGH_MEMORY];
1405
1406         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1407
1408         if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1409                 NODEMASK_FREE(nodes_allowed);
1410
1411         return len;
1412 out:
1413         NODEMASK_FREE(nodes_allowed);
1414         return err;
1415 }
1416
1417 static ssize_t nr_hugepages_show(struct kobject *kobj,
1418                                        struct kobj_attribute *attr, char *buf)
1419 {
1420         return nr_hugepages_show_common(kobj, attr, buf);
1421 }
1422
1423 static ssize_t nr_hugepages_store(struct kobject *kobj,
1424                struct kobj_attribute *attr, const char *buf, size_t len)
1425 {
1426         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1427 }
1428 HSTATE_ATTR(nr_hugepages);
1429
1430 #ifdef CONFIG_NUMA
1431
1432 /*
1433  * hstate attribute for optionally mempolicy-based constraint on persistent
1434  * huge page alloc/free.
1435  */
1436 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1437                                        struct kobj_attribute *attr, char *buf)
1438 {
1439         return nr_hugepages_show_common(kobj, attr, buf);
1440 }
1441
1442 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1443                struct kobj_attribute *attr, const char *buf, size_t len)
1444 {
1445         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1446 }
1447 HSTATE_ATTR(nr_hugepages_mempolicy);
1448 #endif
1449
1450
1451 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1452                                         struct kobj_attribute *attr, char *buf)
1453 {
1454         struct hstate *h = kobj_to_hstate(kobj, NULL);
1455         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1456 }
1457
1458 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1459                 struct kobj_attribute *attr, const char *buf, size_t count)
1460 {
1461         int err;
1462         unsigned long input;
1463         struct hstate *h = kobj_to_hstate(kobj, NULL);
1464
1465         if (h->order >= MAX_ORDER)
1466                 return -EINVAL;
1467
1468         err = strict_strtoul(buf, 10, &input);
1469         if (err)
1470                 return err;
1471
1472         spin_lock(&hugetlb_lock);
1473         h->nr_overcommit_huge_pages = input;
1474         spin_unlock(&hugetlb_lock);
1475
1476         return count;
1477 }
1478 HSTATE_ATTR(nr_overcommit_hugepages);
1479
1480 static ssize_t free_hugepages_show(struct kobject *kobj,
1481                                         struct kobj_attribute *attr, char *buf)
1482 {
1483         struct hstate *h;
1484         unsigned long free_huge_pages;
1485         int nid;
1486
1487         h = kobj_to_hstate(kobj, &nid);
1488         if (nid == NUMA_NO_NODE)
1489                 free_huge_pages = h->free_huge_pages;
1490         else
1491                 free_huge_pages = h->free_huge_pages_node[nid];
1492
1493         return sprintf(buf, "%lu\n", free_huge_pages);
1494 }
1495 HSTATE_ATTR_RO(free_hugepages);
1496
1497 static ssize_t resv_hugepages_show(struct kobject *kobj,
1498                                         struct kobj_attribute *attr, char *buf)
1499 {
1500         struct hstate *h = kobj_to_hstate(kobj, NULL);
1501         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1502 }
1503 HSTATE_ATTR_RO(resv_hugepages);
1504
1505 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1506                                         struct kobj_attribute *attr, char *buf)
1507 {
1508         struct hstate *h;
1509         unsigned long surplus_huge_pages;
1510         int nid;
1511
1512         h = kobj_to_hstate(kobj, &nid);
1513         if (nid == NUMA_NO_NODE)
1514                 surplus_huge_pages = h->surplus_huge_pages;
1515         else
1516                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1517
1518         return sprintf(buf, "%lu\n", surplus_huge_pages);
1519 }
1520 HSTATE_ATTR_RO(surplus_hugepages);
1521
1522 static struct attribute *hstate_attrs[] = {
1523         &nr_hugepages_attr.attr,
1524         &nr_overcommit_hugepages_attr.attr,
1525         &free_hugepages_attr.attr,
1526         &resv_hugepages_attr.attr,
1527         &surplus_hugepages_attr.attr,
1528 #ifdef CONFIG_NUMA
1529         &nr_hugepages_mempolicy_attr.attr,
1530 #endif
1531         NULL,
1532 };
1533
1534 static struct attribute_group hstate_attr_group = {
1535         .attrs = hstate_attrs,
1536 };
1537
1538 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1539                                     struct kobject **hstate_kobjs,
1540                                     struct attribute_group *hstate_attr_group)
1541 {
1542         int retval;
1543         int hi = h - hstates;
1544
1545         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1546         if (!hstate_kobjs[hi])
1547                 return -ENOMEM;
1548
1549         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1550         if (retval)
1551                 kobject_put(hstate_kobjs[hi]);
1552
1553         return retval;
1554 }
1555
1556 static void __init hugetlb_sysfs_init(void)
1557 {
1558         struct hstate *h;
1559         int err;
1560
1561         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1562         if (!hugepages_kobj)
1563                 return;
1564
1565         for_each_hstate(h) {
1566                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1567                                          hstate_kobjs, &hstate_attr_group);
1568                 if (err)
1569                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1570                                                                 h->name);
1571         }
1572 }
1573
1574 #ifdef CONFIG_NUMA
1575
1576 /*
1577  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1578  * with node sysdevs in node_devices[] using a parallel array.  The array
1579  * index of a node sysdev or _hstate == node id.
1580  * This is here to avoid any static dependency of the node sysdev driver, in
1581  * the base kernel, on the hugetlb module.
1582  */
1583 struct node_hstate {
1584         struct kobject          *hugepages_kobj;
1585         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1586 };
1587 struct node_hstate node_hstates[MAX_NUMNODES];
1588
1589 /*
1590  * A subset of global hstate attributes for node sysdevs
1591  */
1592 static struct attribute *per_node_hstate_attrs[] = {
1593         &nr_hugepages_attr.attr,
1594         &free_hugepages_attr.attr,
1595         &surplus_hugepages_attr.attr,
1596         NULL,
1597 };
1598
1599 static struct attribute_group per_node_hstate_attr_group = {
1600         .attrs = per_node_hstate_attrs,
1601 };
1602
1603 /*
1604  * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1605  * Returns node id via non-NULL nidp.
1606  */
1607 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1608 {
1609         int nid;
1610
1611         for (nid = 0; nid < nr_node_ids; nid++) {
1612                 struct node_hstate *nhs = &node_hstates[nid];
1613                 int i;
1614                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1615                         if (nhs->hstate_kobjs[i] == kobj) {
1616                                 if (nidp)
1617                                         *nidp = nid;
1618                                 return &hstates[i];
1619                         }
1620         }
1621
1622         BUG();
1623         return NULL;
1624 }
1625
1626 /*
1627  * Unregister hstate attributes from a single node sysdev.
1628  * No-op if no hstate attributes attached.
1629  */
1630 void hugetlb_unregister_node(struct node *node)
1631 {
1632         struct hstate *h;
1633         struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1634
1635         if (!nhs->hugepages_kobj)
1636                 return;         /* no hstate attributes */
1637
1638         for_each_hstate(h)
1639                 if (nhs->hstate_kobjs[h - hstates]) {
1640                         kobject_put(nhs->hstate_kobjs[h - hstates]);
1641                         nhs->hstate_kobjs[h - hstates] = NULL;
1642                 }
1643
1644         kobject_put(nhs->hugepages_kobj);
1645         nhs->hugepages_kobj = NULL;
1646 }
1647
1648 /*
1649  * hugetlb module exit:  unregister hstate attributes from node sysdevs
1650  * that have them.
1651  */
1652 static void hugetlb_unregister_all_nodes(void)
1653 {
1654         int nid;
1655
1656         /*
1657          * disable node sysdev registrations.
1658          */
1659         register_hugetlbfs_with_node(NULL, NULL);
1660
1661         /*
1662          * remove hstate attributes from any nodes that have them.
1663          */
1664         for (nid = 0; nid < nr_node_ids; nid++)
1665                 hugetlb_unregister_node(&node_devices[nid]);
1666 }
1667
1668 /*
1669  * Register hstate attributes for a single node sysdev.
1670  * No-op if attributes already registered.
1671  */
1672 void hugetlb_register_node(struct node *node)
1673 {
1674         struct hstate *h;
1675         struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1676         int err;
1677
1678         if (nhs->hugepages_kobj)
1679                 return;         /* already allocated */
1680
1681         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1682                                                         &node->sysdev.kobj);
1683         if (!nhs->hugepages_kobj)
1684                 return;
1685
1686         for_each_hstate(h) {
1687                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1688                                                 nhs->hstate_kobjs,
1689                                                 &per_node_hstate_attr_group);
1690                 if (err) {
1691                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1692                                         " for node %d\n",
1693                                                 h->name, node->sysdev.id);
1694                         hugetlb_unregister_node(node);
1695                         break;
1696                 }
1697         }
1698 }
1699
1700 /*
1701  * hugetlb init time:  register hstate attributes for all registered node
1702  * sysdevs of nodes that have memory.  All on-line nodes should have
1703  * registered their associated sysdev by this time.
1704  */
1705 static void hugetlb_register_all_nodes(void)
1706 {
1707         int nid;
1708
1709         for_each_node_state(nid, N_HIGH_MEMORY) {
1710                 struct node *node = &node_devices[nid];
1711                 if (node->sysdev.id == nid)
1712                         hugetlb_register_node(node);
1713         }
1714
1715         /*
1716          * Let the node sysdev driver know we're here so it can
1717          * [un]register hstate attributes on node hotplug.
1718          */
1719         register_hugetlbfs_with_node(hugetlb_register_node,
1720                                      hugetlb_unregister_node);
1721 }
1722 #else   /* !CONFIG_NUMA */
1723
1724 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1725 {
1726         BUG();
1727         if (nidp)
1728                 *nidp = -1;
1729         return NULL;
1730 }
1731
1732 static void hugetlb_unregister_all_nodes(void) { }
1733
1734 static void hugetlb_register_all_nodes(void) { }
1735
1736 #endif
1737
1738 static void __exit hugetlb_exit(void)
1739 {
1740         struct hstate *h;
1741
1742         hugetlb_unregister_all_nodes();
1743
1744         for_each_hstate(h) {
1745                 kobject_put(hstate_kobjs[h - hstates]);
1746         }
1747
1748         kobject_put(hugepages_kobj);
1749 }
1750 module_exit(hugetlb_exit);
1751
1752 static int __init hugetlb_init(void)
1753 {
1754         /* Some platform decide whether they support huge pages at boot
1755          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1756          * there is no such support
1757          */
1758         if (HPAGE_SHIFT == 0)
1759                 return 0;
1760
1761         if (!size_to_hstate(default_hstate_size)) {
1762                 default_hstate_size = HPAGE_SIZE;
1763                 if (!size_to_hstate(default_hstate_size))
1764                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1765         }
1766         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1767         if (default_hstate_max_huge_pages)
1768                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1769
1770         hugetlb_init_hstates();
1771
1772         gather_bootmem_prealloc();
1773
1774         report_hugepages();
1775
1776         hugetlb_sysfs_init();
1777
1778         hugetlb_register_all_nodes();
1779
1780         return 0;
1781 }
1782 module_init(hugetlb_init);
1783
1784 /* Should be called on processing a hugepagesz=... option */
1785 void __init hugetlb_add_hstate(unsigned order)
1786 {
1787         struct hstate *h;
1788         unsigned long i;
1789
1790         if (size_to_hstate(PAGE_SIZE << order)) {
1791                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1792                 return;
1793         }
1794         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1795         BUG_ON(order == 0);
1796         h = &hstates[max_hstate++];
1797         h->order = order;
1798         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1799         h->nr_huge_pages = 0;
1800         h->free_huge_pages = 0;
1801         for (i = 0; i < MAX_NUMNODES; ++i)
1802                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1803         h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1804         h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1805         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1806                                         huge_page_size(h)/1024);
1807
1808         parsed_hstate = h;
1809 }
1810
1811 static int __init hugetlb_nrpages_setup(char *s)
1812 {
1813         unsigned long *mhp;
1814         static unsigned long *last_mhp;
1815
1816         /*
1817          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1818          * so this hugepages= parameter goes to the "default hstate".
1819          */
1820         if (!max_hstate)
1821                 mhp = &default_hstate_max_huge_pages;
1822         else
1823                 mhp = &parsed_hstate->max_huge_pages;
1824
1825         if (mhp == last_mhp) {
1826                 printk(KERN_WARNING "hugepages= specified twice without "
1827                         "interleaving hugepagesz=, ignoring\n");
1828                 return 1;
1829         }
1830
1831         if (sscanf(s, "%lu", mhp) <= 0)
1832                 *mhp = 0;
1833
1834         /*
1835          * Global state is always initialized later in hugetlb_init.
1836          * But we need to allocate >= MAX_ORDER hstates here early to still
1837          * use the bootmem allocator.
1838          */
1839         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1840                 hugetlb_hstate_alloc_pages(parsed_hstate);
1841
1842         last_mhp = mhp;
1843
1844         return 1;
1845 }
1846 __setup("hugepages=", hugetlb_nrpages_setup);
1847
1848 static int __init hugetlb_default_setup(char *s)
1849 {
1850         default_hstate_size = memparse(s, &s);
1851         return 1;
1852 }
1853 __setup("default_hugepagesz=", hugetlb_default_setup);
1854
1855 static unsigned int cpuset_mems_nr(unsigned int *array)
1856 {
1857         int node;
1858         unsigned int nr = 0;
1859
1860         for_each_node_mask(node, cpuset_current_mems_allowed)
1861                 nr += array[node];
1862
1863         return nr;
1864 }
1865
1866 #ifdef CONFIG_SYSCTL
1867 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1868                          struct ctl_table *table, int write,
1869                          void __user *buffer, size_t *length, loff_t *ppos)
1870 {
1871         struct hstate *h = &default_hstate;
1872         unsigned long tmp;
1873         int ret;
1874
1875         tmp = h->max_huge_pages;
1876
1877         if (write && h->order >= MAX_ORDER)
1878                 return -EINVAL;
1879
1880         table->data = &tmp;
1881         table->maxlen = sizeof(unsigned long);
1882         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1883         if (ret)
1884                 goto out;
1885
1886         if (write) {
1887                 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1888                                                 GFP_KERNEL | __GFP_NORETRY);
1889                 if (!(obey_mempolicy &&
1890                                init_nodemask_of_mempolicy(nodes_allowed))) {
1891                         NODEMASK_FREE(nodes_allowed);
1892                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1893                 }
1894                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1895
1896                 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1897                         NODEMASK_FREE(nodes_allowed);
1898         }
1899 out:
1900         return ret;
1901 }
1902
1903 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1904                           void __user *buffer, size_t *length, loff_t *ppos)
1905 {
1906
1907         return hugetlb_sysctl_handler_common(false, table, write,
1908                                                         buffer, length, ppos);
1909 }
1910
1911 #ifdef CONFIG_NUMA
1912 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1913                           void __user *buffer, size_t *length, loff_t *ppos)
1914 {
1915         return hugetlb_sysctl_handler_common(true, table, write,
1916                                                         buffer, length, ppos);
1917 }
1918 #endif /* CONFIG_NUMA */
1919
1920 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1921                         void __user *buffer,
1922                         size_t *length, loff_t *ppos)
1923 {
1924         proc_dointvec(table, write, buffer, length, ppos);
1925         if (hugepages_treat_as_movable)
1926                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1927         else
1928                 htlb_alloc_mask = GFP_HIGHUSER;
1929         return 0;
1930 }
1931
1932 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1933                         void __user *buffer,
1934                         size_t *length, loff_t *ppos)
1935 {
1936         struct hstate *h = &default_hstate;
1937         unsigned long tmp;
1938         int ret;
1939
1940         tmp = h->nr_overcommit_huge_pages;
1941
1942         if (write && h->order >= MAX_ORDER)
1943                 return -EINVAL;
1944
1945         table->data = &tmp;
1946         table->maxlen = sizeof(unsigned long);
1947         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1948         if (ret)
1949                 goto out;
1950
1951         if (write) {
1952                 spin_lock(&hugetlb_lock);
1953                 h->nr_overcommit_huge_pages = tmp;
1954                 spin_unlock(&hugetlb_lock);
1955         }
1956 out:
1957         return ret;
1958 }
1959
1960 #endif /* CONFIG_SYSCTL */
1961
1962 void hugetlb_report_meminfo(struct seq_file *m)
1963 {
1964         struct hstate *h = &default_hstate;
1965         seq_printf(m,
1966                         "HugePages_Total:   %5lu\n"
1967                         "HugePages_Free:    %5lu\n"
1968                         "HugePages_Rsvd:    %5lu\n"
1969                         "HugePages_Surp:    %5lu\n"
1970                         "Hugepagesize:   %8lu kB\n",
1971                         h->nr_huge_pages,
1972                         h->free_huge_pages,
1973                         h->resv_huge_pages,
1974                         h->surplus_huge_pages,
1975                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1976 }
1977
1978 int hugetlb_report_node_meminfo(int nid, char *buf)
1979 {
1980         struct hstate *h = &default_hstate;
1981         return sprintf(buf,
1982                 "Node %d HugePages_Total: %5u\n"
1983                 "Node %d HugePages_Free:  %5u\n"
1984                 "Node %d HugePages_Surp:  %5u\n",
1985                 nid, h->nr_huge_pages_node[nid],
1986                 nid, h->free_huge_pages_node[nid],
1987                 nid, h->surplus_huge_pages_node[nid]);
1988 }
1989
1990 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1991 unsigned long hugetlb_total_pages(void)
1992 {
1993         struct hstate *h = &default_hstate;
1994         return h->nr_huge_pages * pages_per_huge_page(h);
1995 }
1996
1997 static int hugetlb_acct_memory(struct hstate *h, long delta)
1998 {
1999         int ret = -ENOMEM;
2000
2001         spin_lock(&hugetlb_lock);
2002         /*
2003          * When cpuset is configured, it breaks the strict hugetlb page
2004          * reservation as the accounting is done on a global variable. Such
2005          * reservation is completely rubbish in the presence of cpuset because
2006          * the reservation is not checked against page availability for the
2007          * current cpuset. Application can still potentially OOM'ed by kernel
2008          * with lack of free htlb page in cpuset that the task is in.
2009          * Attempt to enforce strict accounting with cpuset is almost
2010          * impossible (or too ugly) because cpuset is too fluid that
2011          * task or memory node can be dynamically moved between cpusets.
2012          *
2013          * The change of semantics for shared hugetlb mapping with cpuset is
2014          * undesirable. However, in order to preserve some of the semantics,
2015          * we fall back to check against current free page availability as
2016          * a best attempt and hopefully to minimize the impact of changing
2017          * semantics that cpuset has.
2018          */
2019         if (delta > 0) {
2020                 if (gather_surplus_pages(h, delta) < 0)
2021                         goto out;
2022
2023                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2024                         return_unused_surplus_pages(h, delta);
2025                         goto out;
2026                 }
2027         }
2028
2029         ret = 0;
2030         if (delta < 0)
2031                 return_unused_surplus_pages(h, (unsigned long) -delta);
2032
2033 out:
2034         spin_unlock(&hugetlb_lock);
2035         return ret;
2036 }
2037
2038 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2039 {
2040         struct resv_map *reservations = vma_resv_map(vma);
2041
2042         /*
2043          * This new VMA should share its siblings reservation map if present.
2044          * The VMA will only ever have a valid reservation map pointer where
2045          * it is being copied for another still existing VMA.  As that VMA
2046          * has a reference to the reservation map it cannot disappear until
2047          * after this open call completes.  It is therefore safe to take a
2048          * new reference here without additional locking.
2049          */
2050         if (reservations)
2051                 kref_get(&reservations->refs);
2052 }
2053
2054 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2055 {
2056         struct hstate *h = hstate_vma(vma);
2057         struct resv_map *reservations = vma_resv_map(vma);
2058         unsigned long reserve;
2059         unsigned long start;
2060         unsigned long end;
2061
2062         if (reservations) {
2063                 start = vma_hugecache_offset(h, vma, vma->vm_start);
2064                 end = vma_hugecache_offset(h, vma, vma->vm_end);
2065
2066                 reserve = (end - start) -
2067                         region_count(&reservations->regions, start, end);
2068
2069                 kref_put(&reservations->refs, resv_map_release);
2070
2071                 if (reserve) {
2072                         hugetlb_acct_memory(h, -reserve);
2073                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2074                 }
2075         }
2076 }
2077
2078 /*
2079  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2080  * handle_mm_fault() to try to instantiate regular-sized pages in the
2081  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2082  * this far.
2083  */
2084 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2085 {
2086         BUG();
2087         return 0;
2088 }
2089
2090 const struct vm_operations_struct hugetlb_vm_ops = {
2091         .fault = hugetlb_vm_op_fault,
2092         .open = hugetlb_vm_op_open,
2093         .close = hugetlb_vm_op_close,
2094 };
2095
2096 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2097                                 int writable)
2098 {
2099         pte_t entry;
2100
2101         if (writable) {
2102                 entry =
2103                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2104         } else {
2105                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2106         }
2107         entry = pte_mkyoung(entry);
2108         entry = pte_mkhuge(entry);
2109
2110         return entry;
2111 }
2112
2113 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2114                                    unsigned long address, pte_t *ptep)
2115 {
2116         pte_t entry;
2117
2118         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2119         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
2120                 update_mmu_cache(vma, address, ptep);
2121         }
2122 }
2123
2124
2125 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2126                             struct vm_area_struct *vma)
2127 {
2128         pte_t *src_pte, *dst_pte, entry;
2129         struct page *ptepage;
2130         unsigned long addr;
2131         int cow;
2132         struct hstate *h = hstate_vma(vma);
2133         unsigned long sz = huge_page_size(h);
2134
2135         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2136
2137         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2138                 src_pte = huge_pte_offset(src, addr);
2139                 if (!src_pte)
2140                         continue;
2141                 dst_pte = huge_pte_alloc(dst, addr, sz);
2142                 if (!dst_pte)
2143                         goto nomem;
2144
2145                 /* If the pagetables are shared don't copy or take references */
2146                 if (dst_pte == src_pte)
2147                         continue;
2148
2149                 spin_lock(&dst->page_table_lock);
2150                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2151                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2152                         if (cow)
2153                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2154                         entry = huge_ptep_get(src_pte);
2155                         ptepage = pte_page(entry);
2156                         get_page(ptepage);
2157                         page_dup_rmap(ptepage);
2158                         set_huge_pte_at(dst, addr, dst_pte, entry);
2159                 }
2160                 spin_unlock(&src->page_table_lock);
2161                 spin_unlock(&dst->page_table_lock);
2162         }
2163         return 0;
2164
2165 nomem:
2166         return -ENOMEM;
2167 }
2168
2169 static int is_hugetlb_entry_migration(pte_t pte)
2170 {
2171         swp_entry_t swp;
2172
2173         if (huge_pte_none(pte) || pte_present(pte))
2174                 return 0;
2175         swp = pte_to_swp_entry(pte);
2176         if (non_swap_entry(swp) && is_migration_entry(swp)) {
2177                 return 1;
2178         } else
2179                 return 0;
2180 }
2181
2182 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2183 {
2184         swp_entry_t swp;
2185
2186         if (huge_pte_none(pte) || pte_present(pte))
2187                 return 0;
2188         swp = pte_to_swp_entry(pte);
2189         if (non_swap_entry(swp) && is_hwpoison_entry(swp)) {
2190                 return 1;
2191         } else
2192                 return 0;
2193 }
2194
2195 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2196                             unsigned long end, struct page *ref_page)
2197 {
2198         struct mm_struct *mm = vma->vm_mm;
2199         unsigned long address;
2200         pte_t *ptep;
2201         pte_t pte;
2202         struct page *page;
2203         struct page *tmp;
2204         struct hstate *h = hstate_vma(vma);
2205         unsigned long sz = huge_page_size(h);
2206
2207         /*
2208          * A page gathering list, protected by per file i_mmap_mutex. The
2209          * lock is used to avoid list corruption from multiple unmapping
2210          * of the same page since we are using page->lru.
2211          */
2212         LIST_HEAD(page_list);
2213
2214         WARN_ON(!is_vm_hugetlb_page(vma));
2215         BUG_ON(start & ~huge_page_mask(h));
2216         BUG_ON(end & ~huge_page_mask(h));
2217
2218         mmu_notifier_invalidate_range_start(mm, start, end);
2219         spin_lock(&mm->page_table_lock);
2220         for (address = start; address < end; address += sz) {
2221                 ptep = huge_pte_offset(mm, address);
2222                 if (!ptep)
2223                         continue;
2224
2225                 if (huge_pmd_unshare(mm, &address, ptep))
2226                         continue;
2227
2228                 /*
2229                  * If a reference page is supplied, it is because a specific
2230                  * page is being unmapped, not a range. Ensure the page we
2231                  * are about to unmap is the actual page of interest.
2232                  */
2233                 if (ref_page) {
2234                         pte = huge_ptep_get(ptep);
2235                         if (huge_pte_none(pte))
2236                                 continue;
2237                         page = pte_page(pte);
2238                         if (page != ref_page)
2239                                 continue;
2240
2241                         /*
2242                          * Mark the VMA as having unmapped its page so that
2243                          * future faults in this VMA will fail rather than
2244                          * looking like data was lost
2245                          */
2246                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2247                 }
2248
2249                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2250                 if (huge_pte_none(pte))
2251                         continue;
2252
2253                 /*
2254                  * HWPoisoned hugepage is already unmapped and dropped reference
2255                  */
2256                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2257                         continue;
2258
2259                 page = pte_page(pte);
2260                 if (pte_dirty(pte))
2261                         set_page_dirty(page);
2262                 list_add(&page->lru, &page_list);
2263         }
2264         spin_unlock(&mm->page_table_lock);
2265         flush_tlb_range(vma, start, end);
2266         mmu_notifier_invalidate_range_end(mm, start, end);
2267         list_for_each_entry_safe(page, tmp, &page_list, lru) {
2268                 page_remove_rmap(page);
2269                 list_del(&page->lru);
2270                 put_page(page);
2271         }
2272 }
2273
2274 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2275                           unsigned long end, struct page *ref_page)
2276 {
2277         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2278         __unmap_hugepage_range(vma, start, end, ref_page);
2279         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2280 }
2281
2282 /*
2283  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2284  * mappping it owns the reserve page for. The intention is to unmap the page
2285  * from other VMAs and let the children be SIGKILLed if they are faulting the
2286  * same region.
2287  */
2288 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2289                                 struct page *page, unsigned long address)
2290 {
2291         struct hstate *h = hstate_vma(vma);
2292         struct vm_area_struct *iter_vma;
2293         struct address_space *mapping;
2294         struct prio_tree_iter iter;
2295         pgoff_t pgoff;
2296
2297         /*
2298          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2299          * from page cache lookup which is in HPAGE_SIZE units.
2300          */
2301         address = address & huge_page_mask(h);
2302         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2303                 + (vma->vm_pgoff >> PAGE_SHIFT);
2304         mapping = (struct address_space *)page_private(page);
2305
2306         /*
2307          * Take the mapping lock for the duration of the table walk. As
2308          * this mapping should be shared between all the VMAs,
2309          * __unmap_hugepage_range() is called as the lock is already held
2310          */
2311         mutex_lock(&mapping->i_mmap_mutex);
2312         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2313                 /* Do not unmap the current VMA */
2314                 if (iter_vma == vma)
2315                         continue;
2316
2317                 /*
2318                  * Unmap the page from other VMAs without their own reserves.
2319                  * They get marked to be SIGKILLed if they fault in these
2320                  * areas. This is because a future no-page fault on this VMA
2321                  * could insert a zeroed page instead of the data existing
2322                  * from the time of fork. This would look like data corruption
2323                  */
2324                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2325                         __unmap_hugepage_range(iter_vma,
2326                                 address, address + huge_page_size(h),
2327                                 page);
2328         }
2329         mutex_unlock(&mapping->i_mmap_mutex);
2330
2331         return 1;
2332 }
2333
2334 /*
2335  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2336  */
2337 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2338                         unsigned long address, pte_t *ptep, pte_t pte,
2339                         struct page *pagecache_page)
2340 {
2341         struct hstate *h = hstate_vma(vma);
2342         struct page *old_page, *new_page;
2343         int avoidcopy;
2344         int outside_reserve = 0;
2345
2346         old_page = pte_page(pte);
2347
2348 retry_avoidcopy:
2349         /* If no-one else is actually using this page, avoid the copy
2350          * and just make the page writable */
2351         avoidcopy = (page_mapcount(old_page) == 1);
2352         if (avoidcopy) {
2353                 if (PageAnon(old_page))
2354                         page_move_anon_rmap(old_page, vma, address);
2355                 set_huge_ptep_writable(vma, address, ptep);
2356                 return 0;
2357         }
2358
2359         /*
2360          * If the process that created a MAP_PRIVATE mapping is about to
2361          * perform a COW due to a shared page count, attempt to satisfy
2362          * the allocation without using the existing reserves. The pagecache
2363          * page is used to determine if the reserve at this address was
2364          * consumed or not. If reserves were used, a partial faulted mapping
2365          * at the time of fork() could consume its reserves on COW instead
2366          * of the full address range.
2367          */
2368         if (!(vma->vm_flags & VM_MAYSHARE) &&
2369                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2370                         old_page != pagecache_page)
2371                 outside_reserve = 1;
2372
2373         page_cache_get(old_page);
2374
2375         /* Drop page_table_lock as buddy allocator may be called */
2376         spin_unlock(&mm->page_table_lock);
2377         new_page = alloc_huge_page(vma, address, outside_reserve);
2378
2379         if (IS_ERR(new_page)) {
2380                 page_cache_release(old_page);
2381
2382                 /*
2383                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2384                  * it is due to references held by a child and an insufficient
2385                  * huge page pool. To guarantee the original mappers
2386                  * reliability, unmap the page from child processes. The child
2387                  * may get SIGKILLed if it later faults.
2388                  */
2389                 if (outside_reserve) {
2390                         BUG_ON(huge_pte_none(pte));
2391                         if (unmap_ref_private(mm, vma, old_page, address)) {
2392                                 BUG_ON(page_count(old_page) != 1);
2393                                 BUG_ON(huge_pte_none(pte));
2394                                 spin_lock(&mm->page_table_lock);
2395                                 goto retry_avoidcopy;
2396                         }
2397                         WARN_ON_ONCE(1);
2398                 }
2399
2400                 /* Caller expects lock to be held */
2401                 spin_lock(&mm->page_table_lock);
2402                 return -PTR_ERR(new_page);
2403         }
2404
2405         /*
2406          * When the original hugepage is shared one, it does not have
2407          * anon_vma prepared.
2408          */
2409         if (unlikely(anon_vma_prepare(vma))) {
2410                 /* Caller expects lock to be held */
2411                 spin_lock(&mm->page_table_lock);
2412                 return VM_FAULT_OOM;
2413         }
2414
2415         copy_user_huge_page(new_page, old_page, address, vma,
2416                             pages_per_huge_page(h));
2417         __SetPageUptodate(new_page);
2418
2419         /*
2420          * Retake the page_table_lock to check for racing updates
2421          * before the page tables are altered
2422          */
2423         spin_lock(&mm->page_table_lock);
2424         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2425         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2426                 /* Break COW */
2427                 mmu_notifier_invalidate_range_start(mm,
2428                         address & huge_page_mask(h),
2429                         (address & huge_page_mask(h)) + huge_page_size(h));
2430                 huge_ptep_clear_flush(vma, address, ptep);
2431                 set_huge_pte_at(mm, address, ptep,
2432                                 make_huge_pte(vma, new_page, 1));
2433                 page_remove_rmap(old_page);
2434                 hugepage_add_new_anon_rmap(new_page, vma, address);
2435                 /* Make the old page be freed below */
2436                 new_page = old_page;
2437                 mmu_notifier_invalidate_range_end(mm,
2438                         address & huge_page_mask(h),
2439                         (address & huge_page_mask(h)) + huge_page_size(h));
2440         }
2441         page_cache_release(new_page);
2442         page_cache_release(old_page);
2443         return 0;
2444 }
2445
2446 /* Return the pagecache page at a given address within a VMA */
2447 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2448                         struct vm_area_struct *vma, unsigned long address)
2449 {
2450         struct address_space *mapping;
2451         pgoff_t idx;
2452
2453         mapping = vma->vm_file->f_mapping;
2454         idx = vma_hugecache_offset(h, vma, address);
2455
2456         return find_lock_page(mapping, idx);
2457 }
2458
2459 /*
2460  * Return whether there is a pagecache page to back given address within VMA.
2461  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2462  */
2463 static bool hugetlbfs_pagecache_present(struct hstate *h,
2464                         struct vm_area_struct *vma, unsigned long address)
2465 {
2466         struct address_space *mapping;
2467         pgoff_t idx;
2468         struct page *page;
2469
2470         mapping = vma->vm_file->f_mapping;
2471         idx = vma_hugecache_offset(h, vma, address);
2472
2473         page = find_get_page(mapping, idx);
2474         if (page)
2475                 put_page(page);
2476         return page != NULL;
2477 }
2478
2479 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2480                         unsigned long address, pte_t *ptep, unsigned int flags)
2481 {
2482         struct hstate *h = hstate_vma(vma);
2483         int ret = VM_FAULT_SIGBUS;
2484         pgoff_t idx;
2485         unsigned long size;
2486         struct page *page;
2487         struct address_space *mapping;
2488         pte_t new_pte;
2489
2490         /*
2491          * Currently, we are forced to kill the process in the event the
2492          * original mapper has unmapped pages from the child due to a failed
2493          * COW. Warn that such a situation has occurred as it may not be obvious
2494          */
2495         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2496                 printk(KERN_WARNING
2497                         "PID %d killed due to inadequate hugepage pool\n",
2498                         current->pid);
2499                 return ret;
2500         }
2501
2502         mapping = vma->vm_file->f_mapping;
2503         idx = vma_hugecache_offset(h, vma, address);
2504
2505         /*
2506          * Use page lock to guard against racing truncation
2507          * before we get page_table_lock.
2508          */
2509 retry:
2510         page = find_lock_page(mapping, idx);
2511         if (!page) {
2512                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2513                 if (idx >= size)
2514                         goto out;
2515                 page = alloc_huge_page(vma, address, 0);
2516                 if (IS_ERR(page)) {
2517                         ret = -PTR_ERR(page);
2518                         goto out;
2519                 }
2520                 clear_huge_page(page, address, pages_per_huge_page(h));
2521                 __SetPageUptodate(page);
2522
2523                 if (vma->vm_flags & VM_MAYSHARE) {
2524                         int err;
2525                         struct inode *inode = mapping->host;
2526
2527                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2528                         if (err) {
2529                                 put_page(page);
2530                                 if (err == -EEXIST)
2531                                         goto retry;
2532                                 goto out;
2533                         }
2534
2535                         spin_lock(&inode->i_lock);
2536                         inode->i_blocks += blocks_per_huge_page(h);
2537                         spin_unlock(&inode->i_lock);
2538                         page_dup_rmap(page);
2539                 } else {
2540                         lock_page(page);
2541                         if (unlikely(anon_vma_prepare(vma))) {
2542                                 ret = VM_FAULT_OOM;
2543                                 goto backout_unlocked;
2544                         }
2545                         hugepage_add_new_anon_rmap(page, vma, address);
2546                 }
2547         } else {
2548                 /*
2549                  * If memory error occurs between mmap() and fault, some process
2550                  * don't have hwpoisoned swap entry for errored virtual address.
2551                  * So we need to block hugepage fault by PG_hwpoison bit check.
2552                  */
2553                 if (unlikely(PageHWPoison(page))) {
2554                         ret = VM_FAULT_HWPOISON | 
2555                               VM_FAULT_SET_HINDEX(h - hstates);
2556                         goto backout_unlocked;
2557                 }
2558                 page_dup_rmap(page);
2559         }
2560
2561         /*
2562          * If we are going to COW a private mapping later, we examine the
2563          * pending reservations for this page now. This will ensure that
2564          * any allocations necessary to record that reservation occur outside
2565          * the spinlock.
2566          */
2567         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2568                 if (vma_needs_reservation(h, vma, address) < 0) {
2569                         ret = VM_FAULT_OOM;
2570                         goto backout_unlocked;
2571                 }
2572
2573         spin_lock(&mm->page_table_lock);
2574         size = i_size_read(mapping->host) >> huge_page_shift(h);
2575         if (idx >= size)
2576                 goto backout;
2577
2578         ret = 0;
2579         if (!huge_pte_none(huge_ptep_get(ptep)))
2580                 goto backout;
2581
2582         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2583                                 && (vma->vm_flags & VM_SHARED)));
2584         set_huge_pte_at(mm, address, ptep, new_pte);
2585
2586         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2587                 /* Optimization, do the COW without a second fault */
2588                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2589         }
2590
2591         spin_unlock(&mm->page_table_lock);
2592         unlock_page(page);
2593 out:
2594         return ret;
2595
2596 backout:
2597         spin_unlock(&mm->page_table_lock);
2598 backout_unlocked:
2599         unlock_page(page);
2600         put_page(page);
2601         goto out;
2602 }
2603
2604 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2605                         unsigned long address, unsigned int flags)
2606 {
2607         pte_t *ptep;
2608         pte_t entry;
2609         int ret;
2610         struct page *page = NULL;
2611         struct page *pagecache_page = NULL;
2612         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2613         struct hstate *h = hstate_vma(vma);
2614
2615         ptep = huge_pte_offset(mm, address);
2616         if (ptep) {
2617                 entry = huge_ptep_get(ptep);
2618                 if (unlikely(is_hugetlb_entry_migration(entry))) {
2619                         migration_entry_wait(mm, (pmd_t *)ptep, address);
2620                         return 0;
2621                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2622                         return VM_FAULT_HWPOISON_LARGE | 
2623                                VM_FAULT_SET_HINDEX(h - hstates);
2624         }
2625
2626         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2627         if (!ptep)
2628                 return VM_FAULT_OOM;
2629
2630         /*
2631          * Serialize hugepage allocation and instantiation, so that we don't
2632          * get spurious allocation failures if two CPUs race to instantiate
2633          * the same page in the page cache.
2634          */
2635         mutex_lock(&hugetlb_instantiation_mutex);
2636         entry = huge_ptep_get(ptep);
2637         if (huge_pte_none(entry)) {
2638                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2639                 goto out_mutex;
2640         }
2641
2642         ret = 0;
2643
2644         /*
2645          * If we are going to COW the mapping later, we examine the pending
2646          * reservations for this page now. This will ensure that any
2647          * allocations necessary to record that reservation occur outside the
2648          * spinlock. For private mappings, we also lookup the pagecache
2649          * page now as it is used to determine if a reservation has been
2650          * consumed.
2651          */
2652         if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2653                 if (vma_needs_reservation(h, vma, address) < 0) {
2654                         ret = VM_FAULT_OOM;
2655                         goto out_mutex;
2656                 }
2657
2658                 if (!(vma->vm_flags & VM_MAYSHARE))
2659                         pagecache_page = hugetlbfs_pagecache_page(h,
2660                                                                 vma, address);
2661         }
2662
2663         /*
2664          * hugetlb_cow() requires page locks of pte_page(entry) and
2665          * pagecache_page, so here we need take the former one
2666          * when page != pagecache_page or !pagecache_page.
2667          * Note that locking order is always pagecache_page -> page,
2668          * so no worry about deadlock.
2669          */
2670         page = pte_page(entry);
2671         if (page != pagecache_page)
2672                 lock_page(page);
2673
2674         spin_lock(&mm->page_table_lock);
2675         /* Check for a racing update before calling hugetlb_cow */
2676         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2677                 goto out_page_table_lock;
2678
2679
2680         if (flags & FAULT_FLAG_WRITE) {
2681                 if (!pte_write(entry)) {
2682                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2683                                                         pagecache_page);
2684                         goto out_page_table_lock;
2685                 }
2686                 entry = pte_mkdirty(entry);
2687         }
2688         entry = pte_mkyoung(entry);
2689         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2690                                                 flags & FAULT_FLAG_WRITE))
2691                 update_mmu_cache(vma, address, ptep);
2692
2693 out_page_table_lock:
2694         spin_unlock(&mm->page_table_lock);
2695
2696         if (pagecache_page) {
2697                 unlock_page(pagecache_page);
2698                 put_page(pagecache_page);
2699         }
2700         if (page != pagecache_page)
2701                 unlock_page(page);
2702
2703 out_mutex:
2704         mutex_unlock(&hugetlb_instantiation_mutex);
2705
2706         return ret;
2707 }
2708
2709 /* Can be overriden by architectures */
2710 __attribute__((weak)) struct page *
2711 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2712                pud_t *pud, int write)
2713 {
2714         BUG();
2715         return NULL;
2716 }
2717
2718 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2719                         struct page **pages, struct vm_area_struct **vmas,
2720                         unsigned long *position, int *length, int i,
2721                         unsigned int flags)
2722 {
2723         unsigned long pfn_offset;
2724         unsigned long vaddr = *position;
2725         int remainder = *length;
2726         struct hstate *h = hstate_vma(vma);
2727
2728         spin_lock(&mm->page_table_lock);
2729         while (vaddr < vma->vm_end && remainder) {
2730                 pte_t *pte;
2731                 int absent;
2732                 struct page *page;
2733
2734                 /*
2735                  * Some archs (sparc64, sh*) have multiple pte_ts to
2736                  * each hugepage.  We have to make sure we get the
2737                  * first, for the page indexing below to work.
2738                  */
2739                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2740                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2741
2742                 /*
2743                  * When coredumping, it suits get_dump_page if we just return
2744                  * an error where there's an empty slot with no huge pagecache
2745                  * to back it.  This way, we avoid allocating a hugepage, and
2746                  * the sparse dumpfile avoids allocating disk blocks, but its
2747                  * huge holes still show up with zeroes where they need to be.
2748                  */
2749                 if (absent && (flags & FOLL_DUMP) &&
2750                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2751                         remainder = 0;
2752                         break;
2753                 }
2754
2755                 if (absent ||
2756                     ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2757                         int ret;
2758
2759                         spin_unlock(&mm->page_table_lock);
2760                         ret = hugetlb_fault(mm, vma, vaddr,
2761                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2762                         spin_lock(&mm->page_table_lock);
2763                         if (!(ret & VM_FAULT_ERROR))
2764                                 continue;
2765
2766                         remainder = 0;
2767                         break;
2768                 }
2769
2770                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2771                 page = pte_page(huge_ptep_get(pte));
2772 same_page:
2773                 if (pages) {
2774                         pages[i] = mem_map_offset(page, pfn_offset);
2775                         get_page(pages[i]);
2776                 }
2777
2778                 if (vmas)
2779                         vmas[i] = vma;
2780
2781                 vaddr += PAGE_SIZE;
2782                 ++pfn_offset;
2783                 --remainder;
2784                 ++i;
2785                 if (vaddr < vma->vm_end && remainder &&
2786                                 pfn_offset < pages_per_huge_page(h)) {
2787                         /*
2788                          * We use pfn_offset to avoid touching the pageframes
2789                          * of this compound page.
2790                          */
2791                         goto same_page;
2792                 }
2793         }
2794         spin_unlock(&mm->page_table_lock);
2795         *length = remainder;
2796         *position = vaddr;
2797
2798         return i ? i : -EFAULT;
2799 }
2800
2801 void hugetlb_change_protection(struct vm_area_struct *vma,
2802                 unsigned long address, unsigned long end, pgprot_t newprot)
2803 {
2804         struct mm_struct *mm = vma->vm_mm;
2805         unsigned long start = address;
2806         pte_t *ptep;
2807         pte_t pte;
2808         struct hstate *h = hstate_vma(vma);
2809
2810         BUG_ON(address >= end);
2811         flush_cache_range(vma, address, end);
2812
2813         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2814         spin_lock(&mm->page_table_lock);
2815         for (; address < end; address += huge_page_size(h)) {
2816                 ptep = huge_pte_offset(mm, address);
2817                 if (!ptep)
2818                         continue;
2819                 if (huge_pmd_unshare(mm, &address, ptep))
2820                         continue;
2821                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2822                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2823                         pte = pte_mkhuge(pte_modify(pte, newprot));
2824                         set_huge_pte_at(mm, address, ptep, pte);
2825                 }
2826         }
2827         spin_unlock(&mm->page_table_lock);
2828         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2829
2830         flush_tlb_range(vma, start, end);
2831 }
2832
2833 int hugetlb_reserve_pages(struct inode *inode,
2834                                         long from, long to,
2835                                         struct vm_area_struct *vma,
2836                                         vm_flags_t vm_flags)
2837 {
2838         long ret, chg;
2839         struct hstate *h = hstate_inode(inode);
2840
2841         /*
2842          * Only apply hugepage reservation if asked. At fault time, an
2843          * attempt will be made for VM_NORESERVE to allocate a page
2844          * and filesystem quota without using reserves
2845          */
2846         if (vm_flags & VM_NORESERVE)
2847                 return 0;
2848
2849         /*
2850          * Shared mappings base their reservation on the number of pages that
2851          * are already allocated on behalf of the file. Private mappings need
2852          * to reserve the full area even if read-only as mprotect() may be
2853          * called to make the mapping read-write. Assume !vma is a shm mapping
2854          */
2855         if (!vma || vma->vm_flags & VM_MAYSHARE)
2856                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2857         else {
2858                 struct resv_map *resv_map = resv_map_alloc();
2859                 if (!resv_map)
2860                         return -ENOMEM;
2861
2862                 chg = to - from;
2863
2864                 set_vma_resv_map(vma, resv_map);
2865                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2866         }
2867
2868         if (chg < 0)
2869                 return chg;
2870
2871         /* There must be enough filesystem quota for the mapping */
2872         if (hugetlb_get_quota(inode->i_mapping, chg))
2873                 return -ENOSPC;
2874
2875         /*
2876          * Check enough hugepages are available for the reservation.
2877          * Hand back the quota if there are not
2878          */
2879         ret = hugetlb_acct_memory(h, chg);
2880         if (ret < 0) {
2881                 hugetlb_put_quota(inode->i_mapping, chg);
2882                 return ret;
2883         }
2884
2885         /*
2886          * Account for the reservations made. Shared mappings record regions
2887          * that have reservations as they are shared by multiple VMAs.
2888          * When the last VMA disappears, the region map says how much
2889          * the reservation was and the page cache tells how much of
2890          * the reservation was consumed. Private mappings are per-VMA and
2891          * only the consumed reservations are tracked. When the VMA
2892          * disappears, the original reservation is the VMA size and the
2893          * consumed reservations are stored in the map. Hence, nothing
2894          * else has to be done for private mappings here
2895          */
2896         if (!vma || vma->vm_flags & VM_MAYSHARE)
2897                 region_add(&inode->i_mapping->private_list, from, to);
2898         return 0;
2899 }
2900
2901 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2902 {
2903         struct hstate *h = hstate_inode(inode);
2904         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2905
2906         spin_lock(&inode->i_lock);
2907         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2908         spin_unlock(&inode->i_lock);
2909
2910         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2911         hugetlb_acct_memory(h, -(chg - freed));
2912 }
2913
2914 #ifdef CONFIG_MEMORY_FAILURE
2915
2916 /* Should be called in hugetlb_lock */
2917 static int is_hugepage_on_freelist(struct page *hpage)
2918 {
2919         struct page *page;
2920         struct page *tmp;
2921         struct hstate *h = page_hstate(hpage);
2922         int nid = page_to_nid(hpage);
2923
2924         list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
2925                 if (page == hpage)
2926                         return 1;
2927         return 0;
2928 }
2929
2930 /*
2931  * This function is called from memory failure code.
2932  * Assume the caller holds page lock of the head page.
2933  */
2934 int dequeue_hwpoisoned_huge_page(struct page *hpage)
2935 {
2936         struct hstate *h = page_hstate(hpage);
2937         int nid = page_to_nid(hpage);
2938         int ret = -EBUSY;
2939
2940         spin_lock(&hugetlb_lock);
2941         if (is_hugepage_on_freelist(hpage)) {
2942                 list_del(&hpage->lru);
2943                 set_page_refcounted(hpage);
2944                 h->free_huge_pages--;
2945                 h->free_huge_pages_node[nid]--;
2946                 ret = 0;
2947         }
2948         spin_unlock(&hugetlb_lock);
2949         return ret;
2950 }
2951 #endif