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