Revert duplicate "mm/hugetlb.c must #include <asm/io.h>"
[linux-2.6.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) William Irwin, April 2004
4  */
5 #include <linux/gfp.h>
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.h>
9 #include <linux/mm.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
21 #include <asm/page.h>
22 #include <asm/pgtable.h>
23 #include <asm/io.h>
24
25 #include <linux/hugetlb.h>
26 #include "internal.h"
27
28 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
29 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
30 unsigned long hugepages_treat_as_movable;
31
32 static int max_hstate;
33 unsigned int default_hstate_idx;
34 struct hstate hstates[HUGE_MAX_HSTATE];
35
36 __initdata LIST_HEAD(huge_boot_pages);
37
38 /* for command line parsing */
39 static struct hstate * __initdata parsed_hstate;
40 static unsigned long __initdata default_hstate_max_huge_pages;
41 static unsigned long __initdata default_hstate_size;
42
43 #define for_each_hstate(h) \
44         for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
45
46 /*
47  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
48  */
49 static DEFINE_SPINLOCK(hugetlb_lock);
50
51 /*
52  * Region tracking -- allows tracking of reservations and instantiated pages
53  *                    across the pages in a mapping.
54  *
55  * The region data structures are protected by a combination of the mmap_sem
56  * and the hugetlb_instantion_mutex.  To access or modify a region the caller
57  * must either hold the mmap_sem for write, or the mmap_sem for read and
58  * the hugetlb_instantiation mutex:
59  *
60  *      down_write(&mm->mmap_sem);
61  * or
62  *      down_read(&mm->mmap_sem);
63  *      mutex_lock(&hugetlb_instantiation_mutex);
64  */
65 struct file_region {
66         struct list_head link;
67         long from;
68         long to;
69 };
70
71 static long region_add(struct list_head *head, long f, long t)
72 {
73         struct file_region *rg, *nrg, *trg;
74
75         /* Locate the region we are either in or before. */
76         list_for_each_entry(rg, head, link)
77                 if (f <= rg->to)
78                         break;
79
80         /* Round our left edge to the current segment if it encloses us. */
81         if (f > rg->from)
82                 f = rg->from;
83
84         /* Check for and consume any regions we now overlap with. */
85         nrg = rg;
86         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
87                 if (&rg->link == head)
88                         break;
89                 if (rg->from > t)
90                         break;
91
92                 /* If this area reaches higher then extend our area to
93                  * include it completely.  If this is not the first area
94                  * which we intend to reuse, free it. */
95                 if (rg->to > t)
96                         t = rg->to;
97                 if (rg != nrg) {
98                         list_del(&rg->link);
99                         kfree(rg);
100                 }
101         }
102         nrg->from = f;
103         nrg->to = t;
104         return 0;
105 }
106
107 static long region_chg(struct list_head *head, long f, long t)
108 {
109         struct file_region *rg, *nrg;
110         long chg = 0;
111
112         /* Locate the region we are before or in. */
113         list_for_each_entry(rg, head, link)
114                 if (f <= rg->to)
115                         break;
116
117         /* If we are below the current region then a new region is required.
118          * Subtle, allocate a new region at the position but make it zero
119          * size such that we can guarantee to record the reservation. */
120         if (&rg->link == head || t < rg->from) {
121                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
122                 if (!nrg)
123                         return -ENOMEM;
124                 nrg->from = f;
125                 nrg->to   = f;
126                 INIT_LIST_HEAD(&nrg->link);
127                 list_add(&nrg->link, rg->link.prev);
128
129                 return t - f;
130         }
131
132         /* Round our left edge to the current segment if it encloses us. */
133         if (f > rg->from)
134                 f = rg->from;
135         chg = t - f;
136
137         /* Check for and consume any regions we now overlap with. */
138         list_for_each_entry(rg, rg->link.prev, link) {
139                 if (&rg->link == head)
140                         break;
141                 if (rg->from > t)
142                         return chg;
143
144                 /* We overlap with this area, if it extends futher than
145                  * us then we must extend ourselves.  Account for its
146                  * existing reservation. */
147                 if (rg->to > t) {
148                         chg += rg->to - t;
149                         t = rg->to;
150                 }
151                 chg -= rg->to - rg->from;
152         }
153         return chg;
154 }
155
156 static long region_truncate(struct list_head *head, long end)
157 {
158         struct file_region *rg, *trg;
159         long chg = 0;
160
161         /* Locate the region we are either in or before. */
162         list_for_each_entry(rg, head, link)
163                 if (end <= rg->to)
164                         break;
165         if (&rg->link == head)
166                 return 0;
167
168         /* If we are in the middle of a region then adjust it. */
169         if (end > rg->from) {
170                 chg = rg->to - end;
171                 rg->to = end;
172                 rg = list_entry(rg->link.next, typeof(*rg), link);
173         }
174
175         /* Drop any remaining regions. */
176         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
177                 if (&rg->link == head)
178                         break;
179                 chg += rg->to - rg->from;
180                 list_del(&rg->link);
181                 kfree(rg);
182         }
183         return chg;
184 }
185
186 static long region_count(struct list_head *head, long f, long t)
187 {
188         struct file_region *rg;
189         long chg = 0;
190
191         /* Locate each segment we overlap with, and count that overlap. */
192         list_for_each_entry(rg, head, link) {
193                 int seg_from;
194                 int seg_to;
195
196                 if (rg->to <= f)
197                         continue;
198                 if (rg->from >= t)
199                         break;
200
201                 seg_from = max(rg->from, f);
202                 seg_to = min(rg->to, t);
203
204                 chg += seg_to - seg_from;
205         }
206
207         return chg;
208 }
209
210 /*
211  * Convert the address within this vma to the page offset within
212  * the mapping, in pagecache page units; huge pages here.
213  */
214 static pgoff_t vma_hugecache_offset(struct hstate *h,
215                         struct vm_area_struct *vma, unsigned long address)
216 {
217         return ((address - vma->vm_start) >> huge_page_shift(h)) +
218                         (vma->vm_pgoff >> huge_page_order(h));
219 }
220
221 /*
222  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
223  * bits of the reservation map pointer, which are always clear due to
224  * alignment.
225  */
226 #define HPAGE_RESV_OWNER    (1UL << 0)
227 #define HPAGE_RESV_UNMAPPED (1UL << 1)
228 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
229
230 /*
231  * These helpers are used to track how many pages are reserved for
232  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
233  * is guaranteed to have their future faults succeed.
234  *
235  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
236  * the reserve counters are updated with the hugetlb_lock held. It is safe
237  * to reset the VMA at fork() time as it is not in use yet and there is no
238  * chance of the global counters getting corrupted as a result of the values.
239  *
240  * The private mapping reservation is represented in a subtly different
241  * manner to a shared mapping.  A shared mapping has a region map associated
242  * with the underlying file, this region map represents the backing file
243  * pages which have ever had a reservation assigned which this persists even
244  * after the page is instantiated.  A private mapping has a region map
245  * associated with the original mmap which is attached to all VMAs which
246  * reference it, this region map represents those offsets which have consumed
247  * reservation ie. where pages have been instantiated.
248  */
249 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
250 {
251         return (unsigned long)vma->vm_private_data;
252 }
253
254 static void set_vma_private_data(struct vm_area_struct *vma,
255                                                         unsigned long value)
256 {
257         vma->vm_private_data = (void *)value;
258 }
259
260 struct resv_map {
261         struct kref refs;
262         struct list_head regions;
263 };
264
265 struct resv_map *resv_map_alloc(void)
266 {
267         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
268         if (!resv_map)
269                 return NULL;
270
271         kref_init(&resv_map->refs);
272         INIT_LIST_HEAD(&resv_map->regions);
273
274         return resv_map;
275 }
276
277 void resv_map_release(struct kref *ref)
278 {
279         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
280
281         /* Clear out any active regions before we release the map. */
282         region_truncate(&resv_map->regions, 0);
283         kfree(resv_map);
284 }
285
286 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
287 {
288         VM_BUG_ON(!is_vm_hugetlb_page(vma));
289         if (!(vma->vm_flags & VM_SHARED))
290                 return (struct resv_map *)(get_vma_private_data(vma) &
291                                                         ~HPAGE_RESV_MASK);
292         return 0;
293 }
294
295 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
296 {
297         VM_BUG_ON(!is_vm_hugetlb_page(vma));
298         VM_BUG_ON(vma->vm_flags & VM_SHARED);
299
300         set_vma_private_data(vma, (get_vma_private_data(vma) &
301                                 HPAGE_RESV_MASK) | (unsigned long)map);
302 }
303
304 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
305 {
306         VM_BUG_ON(!is_vm_hugetlb_page(vma));
307         VM_BUG_ON(vma->vm_flags & VM_SHARED);
308
309         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
310 }
311
312 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
313 {
314         VM_BUG_ON(!is_vm_hugetlb_page(vma));
315
316         return (get_vma_private_data(vma) & flag) != 0;
317 }
318
319 /* Decrement the reserved pages in the hugepage pool by one */
320 static void decrement_hugepage_resv_vma(struct hstate *h,
321                         struct vm_area_struct *vma)
322 {
323         if (vma->vm_flags & VM_NORESERVE)
324                 return;
325
326         if (vma->vm_flags & VM_SHARED) {
327                 /* Shared mappings always use reserves */
328                 h->resv_huge_pages--;
329         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
330                 /*
331                  * Only the process that called mmap() has reserves for
332                  * private mappings.
333                  */
334                 h->resv_huge_pages--;
335         }
336 }
337
338 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
339 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
340 {
341         VM_BUG_ON(!is_vm_hugetlb_page(vma));
342         if (!(vma->vm_flags & VM_SHARED))
343                 vma->vm_private_data = (void *)0;
344 }
345
346 /* Returns true if the VMA has associated reserve pages */
347 static int vma_has_reserves(struct vm_area_struct *vma)
348 {
349         if (vma->vm_flags & VM_SHARED)
350                 return 1;
351         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
352                 return 1;
353         return 0;
354 }
355
356 static void clear_huge_page(struct page *page,
357                         unsigned long addr, unsigned long sz)
358 {
359         int i;
360
361         might_sleep();
362         for (i = 0; i < sz/PAGE_SIZE; i++) {
363                 cond_resched();
364                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
365         }
366 }
367
368 static void copy_huge_page(struct page *dst, struct page *src,
369                            unsigned long addr, struct vm_area_struct *vma)
370 {
371         int i;
372         struct hstate *h = hstate_vma(vma);
373
374         might_sleep();
375         for (i = 0; i < pages_per_huge_page(h); i++) {
376                 cond_resched();
377                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
378         }
379 }
380
381 static void enqueue_huge_page(struct hstate *h, struct page *page)
382 {
383         int nid = page_to_nid(page);
384         list_add(&page->lru, &h->hugepage_freelists[nid]);
385         h->free_huge_pages++;
386         h->free_huge_pages_node[nid]++;
387 }
388
389 static struct page *dequeue_huge_page(struct hstate *h)
390 {
391         int nid;
392         struct page *page = NULL;
393
394         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
395                 if (!list_empty(&h->hugepage_freelists[nid])) {
396                         page = list_entry(h->hugepage_freelists[nid].next,
397                                           struct page, lru);
398                         list_del(&page->lru);
399                         h->free_huge_pages--;
400                         h->free_huge_pages_node[nid]--;
401                         break;
402                 }
403         }
404         return page;
405 }
406
407 static struct page *dequeue_huge_page_vma(struct hstate *h,
408                                 struct vm_area_struct *vma,
409                                 unsigned long address, int avoid_reserve)
410 {
411         int nid;
412         struct page *page = NULL;
413         struct mempolicy *mpol;
414         nodemask_t *nodemask;
415         struct zonelist *zonelist = huge_zonelist(vma, address,
416                                         htlb_alloc_mask, &mpol, &nodemask);
417         struct zone *zone;
418         struct zoneref *z;
419
420         /*
421          * A child process with MAP_PRIVATE mappings created by their parent
422          * have no page reserves. This check ensures that reservations are
423          * not "stolen". The child may still get SIGKILLed
424          */
425         if (!vma_has_reserves(vma) &&
426                         h->free_huge_pages - h->resv_huge_pages == 0)
427                 return NULL;
428
429         /* If reserves cannot be used, ensure enough pages are in the pool */
430         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
431                 return NULL;
432
433         for_each_zone_zonelist_nodemask(zone, z, zonelist,
434                                                 MAX_NR_ZONES - 1, nodemask) {
435                 nid = zone_to_nid(zone);
436                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
437                     !list_empty(&h->hugepage_freelists[nid])) {
438                         page = list_entry(h->hugepage_freelists[nid].next,
439                                           struct page, lru);
440                         list_del(&page->lru);
441                         h->free_huge_pages--;
442                         h->free_huge_pages_node[nid]--;
443
444                         if (!avoid_reserve)
445                                 decrement_hugepage_resv_vma(h, vma);
446
447                         break;
448                 }
449         }
450         mpol_cond_put(mpol);
451         return page;
452 }
453
454 static void update_and_free_page(struct hstate *h, struct page *page)
455 {
456         int i;
457
458         h->nr_huge_pages--;
459         h->nr_huge_pages_node[page_to_nid(page)]--;
460         for (i = 0; i < pages_per_huge_page(h); i++) {
461                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
462                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
463                                 1 << PG_private | 1<< PG_writeback);
464         }
465         set_compound_page_dtor(page, NULL);
466         set_page_refcounted(page);
467         arch_release_hugepage(page);
468         __free_pages(page, huge_page_order(h));
469 }
470
471 struct hstate *size_to_hstate(unsigned long size)
472 {
473         struct hstate *h;
474
475         for_each_hstate(h) {
476                 if (huge_page_size(h) == size)
477                         return h;
478         }
479         return NULL;
480 }
481
482 static void free_huge_page(struct page *page)
483 {
484         /*
485          * Can't pass hstate in here because it is called from the
486          * compound page destructor.
487          */
488         struct hstate *h = page_hstate(page);
489         int nid = page_to_nid(page);
490         struct address_space *mapping;
491
492         mapping = (struct address_space *) page_private(page);
493         set_page_private(page, 0);
494         BUG_ON(page_count(page));
495         INIT_LIST_HEAD(&page->lru);
496
497         spin_lock(&hugetlb_lock);
498         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
499                 update_and_free_page(h, page);
500                 h->surplus_huge_pages--;
501                 h->surplus_huge_pages_node[nid]--;
502         } else {
503                 enqueue_huge_page(h, page);
504         }
505         spin_unlock(&hugetlb_lock);
506         if (mapping)
507                 hugetlb_put_quota(mapping, 1);
508 }
509
510 /*
511  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
512  * balanced by operating on them in a round-robin fashion.
513  * Returns 1 if an adjustment was made.
514  */
515 static int adjust_pool_surplus(struct hstate *h, int delta)
516 {
517         static int prev_nid;
518         int nid = prev_nid;
519         int ret = 0;
520
521         VM_BUG_ON(delta != -1 && delta != 1);
522         do {
523                 nid = next_node(nid, node_online_map);
524                 if (nid == MAX_NUMNODES)
525                         nid = first_node(node_online_map);
526
527                 /* To shrink on this node, there must be a surplus page */
528                 if (delta < 0 && !h->surplus_huge_pages_node[nid])
529                         continue;
530                 /* Surplus cannot exceed the total number of pages */
531                 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
532                                                 h->nr_huge_pages_node[nid])
533                         continue;
534
535                 h->surplus_huge_pages += delta;
536                 h->surplus_huge_pages_node[nid] += delta;
537                 ret = 1;
538                 break;
539         } while (nid != prev_nid);
540
541         prev_nid = nid;
542         return ret;
543 }
544
545 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
546 {
547         set_compound_page_dtor(page, free_huge_page);
548         spin_lock(&hugetlb_lock);
549         h->nr_huge_pages++;
550         h->nr_huge_pages_node[nid]++;
551         spin_unlock(&hugetlb_lock);
552         put_page(page); /* free it into the hugepage allocator */
553 }
554
555 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
556 {
557         struct page *page;
558
559         if (h->order >= MAX_ORDER)
560                 return NULL;
561
562         page = alloc_pages_node(nid,
563                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
564                                                 __GFP_REPEAT|__GFP_NOWARN,
565                 huge_page_order(h));
566         if (page) {
567                 if (arch_prepare_hugepage(page)) {
568                         __free_pages(page, HUGETLB_PAGE_ORDER);
569                         return NULL;
570                 }
571                 prep_new_huge_page(h, page, nid);
572         }
573
574         return page;
575 }
576
577 /*
578  * Use a helper variable to find the next node and then
579  * copy it back to hugetlb_next_nid afterwards:
580  * otherwise there's a window in which a racer might
581  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
582  * But we don't need to use a spin_lock here: it really
583  * doesn't matter if occasionally a racer chooses the
584  * same nid as we do.  Move nid forward in the mask even
585  * if we just successfully allocated a hugepage so that
586  * the next caller gets hugepages on the next node.
587  */
588 static int hstate_next_node(struct hstate *h)
589 {
590         int next_nid;
591         next_nid = next_node(h->hugetlb_next_nid, node_online_map);
592         if (next_nid == MAX_NUMNODES)
593                 next_nid = first_node(node_online_map);
594         h->hugetlb_next_nid = next_nid;
595         return next_nid;
596 }
597
598 static int alloc_fresh_huge_page(struct hstate *h)
599 {
600         struct page *page;
601         int start_nid;
602         int next_nid;
603         int ret = 0;
604
605         start_nid = h->hugetlb_next_nid;
606
607         do {
608                 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
609                 if (page)
610                         ret = 1;
611                 next_nid = hstate_next_node(h);
612         } while (!page && h->hugetlb_next_nid != start_nid);
613
614         if (ret)
615                 count_vm_event(HTLB_BUDDY_PGALLOC);
616         else
617                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
618
619         return ret;
620 }
621
622 static struct page *alloc_buddy_huge_page(struct hstate *h,
623                         struct vm_area_struct *vma, unsigned long address)
624 {
625         struct page *page;
626         unsigned int nid;
627
628         if (h->order >= MAX_ORDER)
629                 return NULL;
630
631         /*
632          * Assume we will successfully allocate the surplus page to
633          * prevent racing processes from causing the surplus to exceed
634          * overcommit
635          *
636          * This however introduces a different race, where a process B
637          * tries to grow the static hugepage pool while alloc_pages() is
638          * called by process A. B will only examine the per-node
639          * counters in determining if surplus huge pages can be
640          * converted to normal huge pages in adjust_pool_surplus(). A
641          * won't be able to increment the per-node counter, until the
642          * lock is dropped by B, but B doesn't drop hugetlb_lock until
643          * no more huge pages can be converted from surplus to normal
644          * state (and doesn't try to convert again). Thus, we have a
645          * case where a surplus huge page exists, the pool is grown, and
646          * the surplus huge page still exists after, even though it
647          * should just have been converted to a normal huge page. This
648          * does not leak memory, though, as the hugepage will be freed
649          * once it is out of use. It also does not allow the counters to
650          * go out of whack in adjust_pool_surplus() as we don't modify
651          * the node values until we've gotten the hugepage and only the
652          * per-node value is checked there.
653          */
654         spin_lock(&hugetlb_lock);
655         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
656                 spin_unlock(&hugetlb_lock);
657                 return NULL;
658         } else {
659                 h->nr_huge_pages++;
660                 h->surplus_huge_pages++;
661         }
662         spin_unlock(&hugetlb_lock);
663
664         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
665                                         __GFP_REPEAT|__GFP_NOWARN,
666                                         huge_page_order(h));
667
668         spin_lock(&hugetlb_lock);
669         if (page) {
670                 /*
671                  * This page is now managed by the hugetlb allocator and has
672                  * no users -- drop the buddy allocator's reference.
673                  */
674                 put_page_testzero(page);
675                 VM_BUG_ON(page_count(page));
676                 nid = page_to_nid(page);
677                 set_compound_page_dtor(page, free_huge_page);
678                 /*
679                  * We incremented the global counters already
680                  */
681                 h->nr_huge_pages_node[nid]++;
682                 h->surplus_huge_pages_node[nid]++;
683                 __count_vm_event(HTLB_BUDDY_PGALLOC);
684         } else {
685                 h->nr_huge_pages--;
686                 h->surplus_huge_pages--;
687                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
688         }
689         spin_unlock(&hugetlb_lock);
690
691         return page;
692 }
693
694 /*
695  * Increase the hugetlb pool such that it can accomodate a reservation
696  * of size 'delta'.
697  */
698 static int gather_surplus_pages(struct hstate *h, int delta)
699 {
700         struct list_head surplus_list;
701         struct page *page, *tmp;
702         int ret, i;
703         int needed, allocated;
704
705         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
706         if (needed <= 0) {
707                 h->resv_huge_pages += delta;
708                 return 0;
709         }
710
711         allocated = 0;
712         INIT_LIST_HEAD(&surplus_list);
713
714         ret = -ENOMEM;
715 retry:
716         spin_unlock(&hugetlb_lock);
717         for (i = 0; i < needed; i++) {
718                 page = alloc_buddy_huge_page(h, NULL, 0);
719                 if (!page) {
720                         /*
721                          * We were not able to allocate enough pages to
722                          * satisfy the entire reservation so we free what
723                          * we've allocated so far.
724                          */
725                         spin_lock(&hugetlb_lock);
726                         needed = 0;
727                         goto free;
728                 }
729
730                 list_add(&page->lru, &surplus_list);
731         }
732         allocated += needed;
733
734         /*
735          * After retaking hugetlb_lock, we need to recalculate 'needed'
736          * because either resv_huge_pages or free_huge_pages may have changed.
737          */
738         spin_lock(&hugetlb_lock);
739         needed = (h->resv_huge_pages + delta) -
740                         (h->free_huge_pages + allocated);
741         if (needed > 0)
742                 goto retry;
743
744         /*
745          * The surplus_list now contains _at_least_ the number of extra pages
746          * needed to accomodate the reservation.  Add the appropriate number
747          * of pages to the hugetlb pool and free the extras back to the buddy
748          * allocator.  Commit the entire reservation here to prevent another
749          * process from stealing the pages as they are added to the pool but
750          * before they are reserved.
751          */
752         needed += allocated;
753         h->resv_huge_pages += delta;
754         ret = 0;
755 free:
756         /* Free the needed pages to the hugetlb pool */
757         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
758                 if ((--needed) < 0)
759                         break;
760                 list_del(&page->lru);
761                 enqueue_huge_page(h, page);
762         }
763
764         /* Free unnecessary surplus pages to the buddy allocator */
765         if (!list_empty(&surplus_list)) {
766                 spin_unlock(&hugetlb_lock);
767                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
768                         list_del(&page->lru);
769                         /*
770                          * The page has a reference count of zero already, so
771                          * call free_huge_page directly instead of using
772                          * put_page.  This must be done with hugetlb_lock
773                          * unlocked which is safe because free_huge_page takes
774                          * hugetlb_lock before deciding how to free the page.
775                          */
776                         free_huge_page(page);
777                 }
778                 spin_lock(&hugetlb_lock);
779         }
780
781         return ret;
782 }
783
784 /*
785  * When releasing a hugetlb pool reservation, any surplus pages that were
786  * allocated to satisfy the reservation must be explicitly freed if they were
787  * never used.
788  */
789 static void return_unused_surplus_pages(struct hstate *h,
790                                         unsigned long unused_resv_pages)
791 {
792         static int nid = -1;
793         struct page *page;
794         unsigned long nr_pages;
795
796         /*
797          * We want to release as many surplus pages as possible, spread
798          * evenly across all nodes. Iterate across all nodes until we
799          * can no longer free unreserved surplus pages. This occurs when
800          * the nodes with surplus pages have no free pages.
801          */
802         unsigned long remaining_iterations = num_online_nodes();
803
804         /* Uncommit the reservation */
805         h->resv_huge_pages -= unused_resv_pages;
806
807         /* Cannot return gigantic pages currently */
808         if (h->order >= MAX_ORDER)
809                 return;
810
811         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
812
813         while (remaining_iterations-- && nr_pages) {
814                 nid = next_node(nid, node_online_map);
815                 if (nid == MAX_NUMNODES)
816                         nid = first_node(node_online_map);
817
818                 if (!h->surplus_huge_pages_node[nid])
819                         continue;
820
821                 if (!list_empty(&h->hugepage_freelists[nid])) {
822                         page = list_entry(h->hugepage_freelists[nid].next,
823                                           struct page, lru);
824                         list_del(&page->lru);
825                         update_and_free_page(h, page);
826                         h->free_huge_pages--;
827                         h->free_huge_pages_node[nid]--;
828                         h->surplus_huge_pages--;
829                         h->surplus_huge_pages_node[nid]--;
830                         nr_pages--;
831                         remaining_iterations = num_online_nodes();
832                 }
833         }
834 }
835
836 /*
837  * Determine if the huge page at addr within the vma has an associated
838  * reservation.  Where it does not we will need to logically increase
839  * reservation and actually increase quota before an allocation can occur.
840  * Where any new reservation would be required the reservation change is
841  * prepared, but not committed.  Once the page has been quota'd allocated
842  * an instantiated the change should be committed via vma_commit_reservation.
843  * No action is required on failure.
844  */
845 static int vma_needs_reservation(struct hstate *h,
846                         struct vm_area_struct *vma, unsigned long addr)
847 {
848         struct address_space *mapping = vma->vm_file->f_mapping;
849         struct inode *inode = mapping->host;
850
851         if (vma->vm_flags & VM_SHARED) {
852                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
853                 return region_chg(&inode->i_mapping->private_list,
854                                                         idx, idx + 1);
855
856         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
857                 return 1;
858
859         } else  {
860                 int err;
861                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
862                 struct resv_map *reservations = vma_resv_map(vma);
863
864                 err = region_chg(&reservations->regions, idx, idx + 1);
865                 if (err < 0)
866                         return err;
867                 return 0;
868         }
869 }
870 static void vma_commit_reservation(struct hstate *h,
871                         struct vm_area_struct *vma, unsigned long addr)
872 {
873         struct address_space *mapping = vma->vm_file->f_mapping;
874         struct inode *inode = mapping->host;
875
876         if (vma->vm_flags & VM_SHARED) {
877                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
878                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
879
880         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
881                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
882                 struct resv_map *reservations = vma_resv_map(vma);
883
884                 /* Mark this page used in the map. */
885                 region_add(&reservations->regions, idx, idx + 1);
886         }
887 }
888
889 static struct page *alloc_huge_page(struct vm_area_struct *vma,
890                                     unsigned long addr, int avoid_reserve)
891 {
892         struct hstate *h = hstate_vma(vma);
893         struct page *page;
894         struct address_space *mapping = vma->vm_file->f_mapping;
895         struct inode *inode = mapping->host;
896         unsigned int chg;
897
898         /*
899          * Processes that did not create the mapping will have no reserves and
900          * will not have accounted against quota. Check that the quota can be
901          * made before satisfying the allocation
902          * MAP_NORESERVE mappings may also need pages and quota allocated
903          * if no reserve mapping overlaps.
904          */
905         chg = vma_needs_reservation(h, vma, addr);
906         if (chg < 0)
907                 return ERR_PTR(chg);
908         if (chg)
909                 if (hugetlb_get_quota(inode->i_mapping, chg))
910                         return ERR_PTR(-ENOSPC);
911
912         spin_lock(&hugetlb_lock);
913         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
914         spin_unlock(&hugetlb_lock);
915
916         if (!page) {
917                 page = alloc_buddy_huge_page(h, vma, addr);
918                 if (!page) {
919                         hugetlb_put_quota(inode->i_mapping, chg);
920                         return ERR_PTR(-VM_FAULT_OOM);
921                 }
922         }
923
924         set_page_refcounted(page);
925         set_page_private(page, (unsigned long) mapping);
926
927         vma_commit_reservation(h, vma, addr);
928
929         return page;
930 }
931
932 __attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
933 {
934         struct huge_bootmem_page *m;
935         int nr_nodes = nodes_weight(node_online_map);
936
937         while (nr_nodes) {
938                 void *addr;
939
940                 addr = __alloc_bootmem_node_nopanic(
941                                 NODE_DATA(h->hugetlb_next_nid),
942                                 huge_page_size(h), huge_page_size(h), 0);
943
944                 if (addr) {
945                         /*
946                          * Use the beginning of the huge page to store the
947                          * huge_bootmem_page struct (until gather_bootmem
948                          * puts them into the mem_map).
949                          */
950                         m = addr;
951                         if (m)
952                                 goto found;
953                 }
954                 hstate_next_node(h);
955                 nr_nodes--;
956         }
957         return 0;
958
959 found:
960         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
961         /* Put them into a private list first because mem_map is not up yet */
962         list_add(&m->list, &huge_boot_pages);
963         m->hstate = h;
964         return 1;
965 }
966
967 /* Put bootmem huge pages into the standard lists after mem_map is up */
968 static void __init gather_bootmem_prealloc(void)
969 {
970         struct huge_bootmem_page *m;
971
972         list_for_each_entry(m, &huge_boot_pages, list) {
973                 struct page *page = virt_to_page(m);
974                 struct hstate *h = m->hstate;
975                 __ClearPageReserved(page);
976                 WARN_ON(page_count(page) != 1);
977                 prep_compound_page(page, h->order);
978                 prep_new_huge_page(h, page, page_to_nid(page));
979         }
980 }
981
982 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
983 {
984         unsigned long i;
985
986         for (i = 0; i < h->max_huge_pages; ++i) {
987                 if (h->order >= MAX_ORDER) {
988                         if (!alloc_bootmem_huge_page(h))
989                                 break;
990                 } else if (!alloc_fresh_huge_page(h))
991                         break;
992         }
993         h->max_huge_pages = i;
994 }
995
996 static void __init hugetlb_init_hstates(void)
997 {
998         struct hstate *h;
999
1000         for_each_hstate(h) {
1001                 /* oversize hugepages were init'ed in early boot */
1002                 if (h->order < MAX_ORDER)
1003                         hugetlb_hstate_alloc_pages(h);
1004         }
1005 }
1006
1007 static char * __init memfmt(char *buf, unsigned long n)
1008 {
1009         if (n >= (1UL << 30))
1010                 sprintf(buf, "%lu GB", n >> 30);
1011         else if (n >= (1UL << 20))
1012                 sprintf(buf, "%lu MB", n >> 20);
1013         else
1014                 sprintf(buf, "%lu KB", n >> 10);
1015         return buf;
1016 }
1017
1018 static void __init report_hugepages(void)
1019 {
1020         struct hstate *h;
1021
1022         for_each_hstate(h) {
1023                 char buf[32];
1024                 printk(KERN_INFO "HugeTLB registered %s page size, "
1025                                  "pre-allocated %ld pages\n",
1026                         memfmt(buf, huge_page_size(h)),
1027                         h->free_huge_pages);
1028         }
1029 }
1030
1031 #ifdef CONFIG_HIGHMEM
1032 static void try_to_free_low(struct hstate *h, unsigned long count)
1033 {
1034         int i;
1035
1036         if (h->order >= MAX_ORDER)
1037                 return;
1038
1039         for (i = 0; i < MAX_NUMNODES; ++i) {
1040                 struct page *page, *next;
1041                 struct list_head *freel = &h->hugepage_freelists[i];
1042                 list_for_each_entry_safe(page, next, freel, lru) {
1043                         if (count >= h->nr_huge_pages)
1044                                 return;
1045                         if (PageHighMem(page))
1046                                 continue;
1047                         list_del(&page->lru);
1048                         update_and_free_page(h, page);
1049                         h->free_huge_pages--;
1050                         h->free_huge_pages_node[page_to_nid(page)]--;
1051                 }
1052         }
1053 }
1054 #else
1055 static inline void try_to_free_low(struct hstate *h, unsigned long count)
1056 {
1057 }
1058 #endif
1059
1060 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1061 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
1062 {
1063         unsigned long min_count, ret;
1064
1065         if (h->order >= MAX_ORDER)
1066                 return h->max_huge_pages;
1067
1068         /*
1069          * Increase the pool size
1070          * First take pages out of surplus state.  Then make up the
1071          * remaining difference by allocating fresh huge pages.
1072          *
1073          * We might race with alloc_buddy_huge_page() here and be unable
1074          * to convert a surplus huge page to a normal huge page. That is
1075          * not critical, though, it just means the overall size of the
1076          * pool might be one hugepage larger than it needs to be, but
1077          * within all the constraints specified by the sysctls.
1078          */
1079         spin_lock(&hugetlb_lock);
1080         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1081                 if (!adjust_pool_surplus(h, -1))
1082                         break;
1083         }
1084
1085         while (count > persistent_huge_pages(h)) {
1086                 /*
1087                  * If this allocation races such that we no longer need the
1088                  * page, free_huge_page will handle it by freeing the page
1089                  * and reducing the surplus.
1090                  */
1091                 spin_unlock(&hugetlb_lock);
1092                 ret = alloc_fresh_huge_page(h);
1093                 spin_lock(&hugetlb_lock);
1094                 if (!ret)
1095                         goto out;
1096
1097         }
1098
1099         /*
1100          * Decrease the pool size
1101          * First return free pages to the buddy allocator (being careful
1102          * to keep enough around to satisfy reservations).  Then place
1103          * pages into surplus state as needed so the pool will shrink
1104          * to the desired size as pages become free.
1105          *
1106          * By placing pages into the surplus state independent of the
1107          * overcommit value, we are allowing the surplus pool size to
1108          * exceed overcommit. There are few sane options here. Since
1109          * alloc_buddy_huge_page() is checking the global counter,
1110          * though, we'll note that we're not allowed to exceed surplus
1111          * and won't grow the pool anywhere else. Not until one of the
1112          * sysctls are changed, or the surplus pages go out of use.
1113          */
1114         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1115         min_count = max(count, min_count);
1116         try_to_free_low(h, min_count);
1117         while (min_count < persistent_huge_pages(h)) {
1118                 struct page *page = dequeue_huge_page(h);
1119                 if (!page)
1120                         break;
1121                 update_and_free_page(h, page);
1122         }
1123         while (count < persistent_huge_pages(h)) {
1124                 if (!adjust_pool_surplus(h, 1))
1125                         break;
1126         }
1127 out:
1128         ret = persistent_huge_pages(h);
1129         spin_unlock(&hugetlb_lock);
1130         return ret;
1131 }
1132
1133 #define HSTATE_ATTR_RO(_name) \
1134         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1135
1136 #define HSTATE_ATTR(_name) \
1137         static struct kobj_attribute _name##_attr = \
1138                 __ATTR(_name, 0644, _name##_show, _name##_store)
1139
1140 static struct kobject *hugepages_kobj;
1141 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1142
1143 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1144 {
1145         int i;
1146         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1147                 if (hstate_kobjs[i] == kobj)
1148                         return &hstates[i];
1149         BUG();
1150         return NULL;
1151 }
1152
1153 static ssize_t nr_hugepages_show(struct kobject *kobj,
1154                                         struct kobj_attribute *attr, char *buf)
1155 {
1156         struct hstate *h = kobj_to_hstate(kobj);
1157         return sprintf(buf, "%lu\n", h->nr_huge_pages);
1158 }
1159 static ssize_t nr_hugepages_store(struct kobject *kobj,
1160                 struct kobj_attribute *attr, const char *buf, size_t count)
1161 {
1162         int err;
1163         unsigned long input;
1164         struct hstate *h = kobj_to_hstate(kobj);
1165
1166         err = strict_strtoul(buf, 10, &input);
1167         if (err)
1168                 return 0;
1169
1170         h->max_huge_pages = set_max_huge_pages(h, input);
1171
1172         return count;
1173 }
1174 HSTATE_ATTR(nr_hugepages);
1175
1176 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1177                                         struct kobj_attribute *attr, char *buf)
1178 {
1179         struct hstate *h = kobj_to_hstate(kobj);
1180         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1181 }
1182 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1183                 struct kobj_attribute *attr, const char *buf, size_t count)
1184 {
1185         int err;
1186         unsigned long input;
1187         struct hstate *h = kobj_to_hstate(kobj);
1188
1189         err = strict_strtoul(buf, 10, &input);
1190         if (err)
1191                 return 0;
1192
1193         spin_lock(&hugetlb_lock);
1194         h->nr_overcommit_huge_pages = input;
1195         spin_unlock(&hugetlb_lock);
1196
1197         return count;
1198 }
1199 HSTATE_ATTR(nr_overcommit_hugepages);
1200
1201 static ssize_t free_hugepages_show(struct kobject *kobj,
1202                                         struct kobj_attribute *attr, char *buf)
1203 {
1204         struct hstate *h = kobj_to_hstate(kobj);
1205         return sprintf(buf, "%lu\n", h->free_huge_pages);
1206 }
1207 HSTATE_ATTR_RO(free_hugepages);
1208
1209 static ssize_t resv_hugepages_show(struct kobject *kobj,
1210                                         struct kobj_attribute *attr, char *buf)
1211 {
1212         struct hstate *h = kobj_to_hstate(kobj);
1213         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1214 }
1215 HSTATE_ATTR_RO(resv_hugepages);
1216
1217 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1218                                         struct kobj_attribute *attr, char *buf)
1219 {
1220         struct hstate *h = kobj_to_hstate(kobj);
1221         return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1222 }
1223 HSTATE_ATTR_RO(surplus_hugepages);
1224
1225 static struct attribute *hstate_attrs[] = {
1226         &nr_hugepages_attr.attr,
1227         &nr_overcommit_hugepages_attr.attr,
1228         &free_hugepages_attr.attr,
1229         &resv_hugepages_attr.attr,
1230         &surplus_hugepages_attr.attr,
1231         NULL,
1232 };
1233
1234 static struct attribute_group hstate_attr_group = {
1235         .attrs = hstate_attrs,
1236 };
1237
1238 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1239 {
1240         int retval;
1241
1242         hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1243                                                         hugepages_kobj);
1244         if (!hstate_kobjs[h - hstates])
1245                 return -ENOMEM;
1246
1247         retval = sysfs_create_group(hstate_kobjs[h - hstates],
1248                                                         &hstate_attr_group);
1249         if (retval)
1250                 kobject_put(hstate_kobjs[h - hstates]);
1251
1252         return retval;
1253 }
1254
1255 static void __init hugetlb_sysfs_init(void)
1256 {
1257         struct hstate *h;
1258         int err;
1259
1260         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1261         if (!hugepages_kobj)
1262                 return;
1263
1264         for_each_hstate(h) {
1265                 err = hugetlb_sysfs_add_hstate(h);
1266                 if (err)
1267                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1268                                                                 h->name);
1269         }
1270 }
1271
1272 static void __exit hugetlb_exit(void)
1273 {
1274         struct hstate *h;
1275
1276         for_each_hstate(h) {
1277                 kobject_put(hstate_kobjs[h - hstates]);
1278         }
1279
1280         kobject_put(hugepages_kobj);
1281 }
1282 module_exit(hugetlb_exit);
1283
1284 static int __init hugetlb_init(void)
1285 {
1286         /* Some platform decide whether they support huge pages at boot
1287          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1288          * there is no such support
1289          */
1290         if (HPAGE_SHIFT == 0)
1291                 return 0;
1292
1293         if (!size_to_hstate(default_hstate_size)) {
1294                 default_hstate_size = HPAGE_SIZE;
1295                 if (!size_to_hstate(default_hstate_size))
1296                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1297         }
1298         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1299         if (default_hstate_max_huge_pages)
1300                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1301
1302         hugetlb_init_hstates();
1303
1304         gather_bootmem_prealloc();
1305
1306         report_hugepages();
1307
1308         hugetlb_sysfs_init();
1309
1310         return 0;
1311 }
1312 module_init(hugetlb_init);
1313
1314 /* Should be called on processing a hugepagesz=... option */
1315 void __init hugetlb_add_hstate(unsigned order)
1316 {
1317         struct hstate *h;
1318         unsigned long i;
1319
1320         if (size_to_hstate(PAGE_SIZE << order)) {
1321                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1322                 return;
1323         }
1324         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1325         BUG_ON(order == 0);
1326         h = &hstates[max_hstate++];
1327         h->order = order;
1328         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1329         h->nr_huge_pages = 0;
1330         h->free_huge_pages = 0;
1331         for (i = 0; i < MAX_NUMNODES; ++i)
1332                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1333         h->hugetlb_next_nid = first_node(node_online_map);
1334         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1335                                         huge_page_size(h)/1024);
1336
1337         parsed_hstate = h;
1338 }
1339
1340 static int __init hugetlb_nrpages_setup(char *s)
1341 {
1342         unsigned long *mhp;
1343         static unsigned long *last_mhp;
1344
1345         /*
1346          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1347          * so this hugepages= parameter goes to the "default hstate".
1348          */
1349         if (!max_hstate)
1350                 mhp = &default_hstate_max_huge_pages;
1351         else
1352                 mhp = &parsed_hstate->max_huge_pages;
1353
1354         if (mhp == last_mhp) {
1355                 printk(KERN_WARNING "hugepages= specified twice without "
1356                         "interleaving hugepagesz=, ignoring\n");
1357                 return 1;
1358         }
1359
1360         if (sscanf(s, "%lu", mhp) <= 0)
1361                 *mhp = 0;
1362
1363         /*
1364          * Global state is always initialized later in hugetlb_init.
1365          * But we need to allocate >= MAX_ORDER hstates here early to still
1366          * use the bootmem allocator.
1367          */
1368         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1369                 hugetlb_hstate_alloc_pages(parsed_hstate);
1370
1371         last_mhp = mhp;
1372
1373         return 1;
1374 }
1375 __setup("hugepages=", hugetlb_nrpages_setup);
1376
1377 static int __init hugetlb_default_setup(char *s)
1378 {
1379         default_hstate_size = memparse(s, &s);
1380         return 1;
1381 }
1382 __setup("default_hugepagesz=", hugetlb_default_setup);
1383
1384 static unsigned int cpuset_mems_nr(unsigned int *array)
1385 {
1386         int node;
1387         unsigned int nr = 0;
1388
1389         for_each_node_mask(node, cpuset_current_mems_allowed)
1390                 nr += array[node];
1391
1392         return nr;
1393 }
1394
1395 #ifdef CONFIG_SYSCTL
1396 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1397                            struct file *file, void __user *buffer,
1398                            size_t *length, loff_t *ppos)
1399 {
1400         struct hstate *h = &default_hstate;
1401         unsigned long tmp;
1402
1403         if (!write)
1404                 tmp = h->max_huge_pages;
1405
1406         table->data = &tmp;
1407         table->maxlen = sizeof(unsigned long);
1408         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1409
1410         if (write)
1411                 h->max_huge_pages = set_max_huge_pages(h, tmp);
1412
1413         return 0;
1414 }
1415
1416 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1417                         struct file *file, void __user *buffer,
1418                         size_t *length, loff_t *ppos)
1419 {
1420         proc_dointvec(table, write, file, buffer, length, ppos);
1421         if (hugepages_treat_as_movable)
1422                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1423         else
1424                 htlb_alloc_mask = GFP_HIGHUSER;
1425         return 0;
1426 }
1427
1428 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1429                         struct file *file, void __user *buffer,
1430                         size_t *length, loff_t *ppos)
1431 {
1432         struct hstate *h = &default_hstate;
1433         unsigned long tmp;
1434
1435         if (!write)
1436                 tmp = h->nr_overcommit_huge_pages;
1437
1438         table->data = &tmp;
1439         table->maxlen = sizeof(unsigned long);
1440         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1441
1442         if (write) {
1443                 spin_lock(&hugetlb_lock);
1444                 h->nr_overcommit_huge_pages = tmp;
1445                 spin_unlock(&hugetlb_lock);
1446         }
1447
1448         return 0;
1449 }
1450
1451 #endif /* CONFIG_SYSCTL */
1452
1453 int hugetlb_report_meminfo(char *buf)
1454 {
1455         struct hstate *h = &default_hstate;
1456         return sprintf(buf,
1457                         "HugePages_Total: %5lu\n"
1458                         "HugePages_Free:  %5lu\n"
1459                         "HugePages_Rsvd:  %5lu\n"
1460                         "HugePages_Surp:  %5lu\n"
1461                         "Hugepagesize:    %5lu kB\n",
1462                         h->nr_huge_pages,
1463                         h->free_huge_pages,
1464                         h->resv_huge_pages,
1465                         h->surplus_huge_pages,
1466                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1467 }
1468
1469 int hugetlb_report_node_meminfo(int nid, char *buf)
1470 {
1471         struct hstate *h = &default_hstate;
1472         return sprintf(buf,
1473                 "Node %d HugePages_Total: %5u\n"
1474                 "Node %d HugePages_Free:  %5u\n"
1475                 "Node %d HugePages_Surp:  %5u\n",
1476                 nid, h->nr_huge_pages_node[nid],
1477                 nid, h->free_huge_pages_node[nid],
1478                 nid, h->surplus_huge_pages_node[nid]);
1479 }
1480
1481 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1482 unsigned long hugetlb_total_pages(void)
1483 {
1484         struct hstate *h = &default_hstate;
1485         return h->nr_huge_pages * pages_per_huge_page(h);
1486 }
1487
1488 static int hugetlb_acct_memory(struct hstate *h, long delta)
1489 {
1490         int ret = -ENOMEM;
1491
1492         spin_lock(&hugetlb_lock);
1493         /*
1494          * When cpuset is configured, it breaks the strict hugetlb page
1495          * reservation as the accounting is done on a global variable. Such
1496          * reservation is completely rubbish in the presence of cpuset because
1497          * the reservation is not checked against page availability for the
1498          * current cpuset. Application can still potentially OOM'ed by kernel
1499          * with lack of free htlb page in cpuset that the task is in.
1500          * Attempt to enforce strict accounting with cpuset is almost
1501          * impossible (or too ugly) because cpuset is too fluid that
1502          * task or memory node can be dynamically moved between cpusets.
1503          *
1504          * The change of semantics for shared hugetlb mapping with cpuset is
1505          * undesirable. However, in order to preserve some of the semantics,
1506          * we fall back to check against current free page availability as
1507          * a best attempt and hopefully to minimize the impact of changing
1508          * semantics that cpuset has.
1509          */
1510         if (delta > 0) {
1511                 if (gather_surplus_pages(h, delta) < 0)
1512                         goto out;
1513
1514                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1515                         return_unused_surplus_pages(h, delta);
1516                         goto out;
1517                 }
1518         }
1519
1520         ret = 0;
1521         if (delta < 0)
1522                 return_unused_surplus_pages(h, (unsigned long) -delta);
1523
1524 out:
1525         spin_unlock(&hugetlb_lock);
1526         return ret;
1527 }
1528
1529 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1530 {
1531         struct resv_map *reservations = vma_resv_map(vma);
1532
1533         /*
1534          * This new VMA should share its siblings reservation map if present.
1535          * The VMA will only ever have a valid reservation map pointer where
1536          * it is being copied for another still existing VMA.  As that VMA
1537          * has a reference to the reservation map it cannot dissappear until
1538          * after this open call completes.  It is therefore safe to take a
1539          * new reference here without additional locking.
1540          */
1541         if (reservations)
1542                 kref_get(&reservations->refs);
1543 }
1544
1545 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1546 {
1547         struct hstate *h = hstate_vma(vma);
1548         struct resv_map *reservations = vma_resv_map(vma);
1549         unsigned long reserve;
1550         unsigned long start;
1551         unsigned long end;
1552
1553         if (reservations) {
1554                 start = vma_hugecache_offset(h, vma, vma->vm_start);
1555                 end = vma_hugecache_offset(h, vma, vma->vm_end);
1556
1557                 reserve = (end - start) -
1558                         region_count(&reservations->regions, start, end);
1559
1560                 kref_put(&reservations->refs, resv_map_release);
1561
1562                 if (reserve) {
1563                         hugetlb_acct_memory(h, -reserve);
1564                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1565                 }
1566         }
1567 }
1568
1569 /*
1570  * We cannot handle pagefaults against hugetlb pages at all.  They cause
1571  * handle_mm_fault() to try to instantiate regular-sized pages in the
1572  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
1573  * this far.
1574  */
1575 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1576 {
1577         BUG();
1578         return 0;
1579 }
1580
1581 struct vm_operations_struct hugetlb_vm_ops = {
1582         .fault = hugetlb_vm_op_fault,
1583         .open = hugetlb_vm_op_open,
1584         .close = hugetlb_vm_op_close,
1585 };
1586
1587 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1588                                 int writable)
1589 {
1590         pte_t entry;
1591
1592         if (writable) {
1593                 entry =
1594                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1595         } else {
1596                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1597         }
1598         entry = pte_mkyoung(entry);
1599         entry = pte_mkhuge(entry);
1600
1601         return entry;
1602 }
1603
1604 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1605                                    unsigned long address, pte_t *ptep)
1606 {
1607         pte_t entry;
1608
1609         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1610         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1611                 update_mmu_cache(vma, address, entry);
1612         }
1613 }
1614
1615
1616 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1617                             struct vm_area_struct *vma)
1618 {
1619         pte_t *src_pte, *dst_pte, entry;
1620         struct page *ptepage;
1621         unsigned long addr;
1622         int cow;
1623         struct hstate *h = hstate_vma(vma);
1624         unsigned long sz = huge_page_size(h);
1625
1626         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1627
1628         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1629                 src_pte = huge_pte_offset(src, addr);
1630                 if (!src_pte)
1631                         continue;
1632                 dst_pte = huge_pte_alloc(dst, addr, sz);
1633                 if (!dst_pte)
1634                         goto nomem;
1635
1636                 /* If the pagetables are shared don't copy or take references */
1637                 if (dst_pte == src_pte)
1638                         continue;
1639
1640                 spin_lock(&dst->page_table_lock);
1641                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1642                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1643                         if (cow)
1644                                 huge_ptep_set_wrprotect(src, addr, src_pte);
1645                         entry = huge_ptep_get(src_pte);
1646                         ptepage = pte_page(entry);
1647                         get_page(ptepage);
1648                         set_huge_pte_at(dst, addr, dst_pte, entry);
1649                 }
1650                 spin_unlock(&src->page_table_lock);
1651                 spin_unlock(&dst->page_table_lock);
1652         }
1653         return 0;
1654
1655 nomem:
1656         return -ENOMEM;
1657 }
1658
1659 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1660                             unsigned long end, struct page *ref_page)
1661 {
1662         struct mm_struct *mm = vma->vm_mm;
1663         unsigned long address;
1664         pte_t *ptep;
1665         pte_t pte;
1666         struct page *page;
1667         struct page *tmp;
1668         struct hstate *h = hstate_vma(vma);
1669         unsigned long sz = huge_page_size(h);
1670
1671         /*
1672          * A page gathering list, protected by per file i_mmap_lock. The
1673          * lock is used to avoid list corruption from multiple unmapping
1674          * of the same page since we are using page->lru.
1675          */
1676         LIST_HEAD(page_list);
1677
1678         WARN_ON(!is_vm_hugetlb_page(vma));
1679         BUG_ON(start & ~huge_page_mask(h));
1680         BUG_ON(end & ~huge_page_mask(h));
1681
1682         mmu_notifier_invalidate_range_start(mm, start, end);
1683         spin_lock(&mm->page_table_lock);
1684         for (address = start; address < end; address += sz) {
1685                 ptep = huge_pte_offset(mm, address);
1686                 if (!ptep)
1687                         continue;
1688
1689                 if (huge_pmd_unshare(mm, &address, ptep))
1690                         continue;
1691
1692                 /*
1693                  * If a reference page is supplied, it is because a specific
1694                  * page is being unmapped, not a range. Ensure the page we
1695                  * are about to unmap is the actual page of interest.
1696                  */
1697                 if (ref_page) {
1698                         pte = huge_ptep_get(ptep);
1699                         if (huge_pte_none(pte))
1700                                 continue;
1701                         page = pte_page(pte);
1702                         if (page != ref_page)
1703                                 continue;
1704
1705                         /*
1706                          * Mark the VMA as having unmapped its page so that
1707                          * future faults in this VMA will fail rather than
1708                          * looking like data was lost
1709                          */
1710                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1711                 }
1712
1713                 pte = huge_ptep_get_and_clear(mm, address, ptep);
1714                 if (huge_pte_none(pte))
1715                         continue;
1716
1717                 page = pte_page(pte);
1718                 if (pte_dirty(pte))
1719                         set_page_dirty(page);
1720                 list_add(&page->lru, &page_list);
1721         }
1722         spin_unlock(&mm->page_table_lock);
1723         flush_tlb_range(vma, start, end);
1724         mmu_notifier_invalidate_range_end(mm, start, end);
1725         list_for_each_entry_safe(page, tmp, &page_list, lru) {
1726                 list_del(&page->lru);
1727                 put_page(page);
1728         }
1729 }
1730
1731 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1732                           unsigned long end, struct page *ref_page)
1733 {
1734         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1735         __unmap_hugepage_range(vma, start, end, ref_page);
1736         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1737 }
1738
1739 /*
1740  * This is called when the original mapper is failing to COW a MAP_PRIVATE
1741  * mappping it owns the reserve page for. The intention is to unmap the page
1742  * from other VMAs and let the children be SIGKILLed if they are faulting the
1743  * same region.
1744  */
1745 int unmap_ref_private(struct mm_struct *mm,
1746                                         struct vm_area_struct *vma,
1747                                         struct page *page,
1748                                         unsigned long address)
1749 {
1750         struct vm_area_struct *iter_vma;
1751         struct address_space *mapping;
1752         struct prio_tree_iter iter;
1753         pgoff_t pgoff;
1754
1755         /*
1756          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1757          * from page cache lookup which is in HPAGE_SIZE units.
1758          */
1759         address = address & huge_page_mask(hstate_vma(vma));
1760         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1761                 + (vma->vm_pgoff >> PAGE_SHIFT);
1762         mapping = (struct address_space *)page_private(page);
1763
1764         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1765                 /* Do not unmap the current VMA */
1766                 if (iter_vma == vma)
1767                         continue;
1768
1769                 /*
1770                  * Unmap the page from other VMAs without their own reserves.
1771                  * They get marked to be SIGKILLed if they fault in these
1772                  * areas. This is because a future no-page fault on this VMA
1773                  * could insert a zeroed page instead of the data existing
1774                  * from the time of fork. This would look like data corruption
1775                  */
1776                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1777                         unmap_hugepage_range(iter_vma,
1778                                 address, address + HPAGE_SIZE,
1779                                 page);
1780         }
1781
1782         return 1;
1783 }
1784
1785 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1786                         unsigned long address, pte_t *ptep, pte_t pte,
1787                         struct page *pagecache_page)
1788 {
1789         struct hstate *h = hstate_vma(vma);
1790         struct page *old_page, *new_page;
1791         int avoidcopy;
1792         int outside_reserve = 0;
1793
1794         old_page = pte_page(pte);
1795
1796 retry_avoidcopy:
1797         /* If no-one else is actually using this page, avoid the copy
1798          * and just make the page writable */
1799         avoidcopy = (page_count(old_page) == 1);
1800         if (avoidcopy) {
1801                 set_huge_ptep_writable(vma, address, ptep);
1802                 return 0;
1803         }
1804
1805         /*
1806          * If the process that created a MAP_PRIVATE mapping is about to
1807          * perform a COW due to a shared page count, attempt to satisfy
1808          * the allocation without using the existing reserves. The pagecache
1809          * page is used to determine if the reserve at this address was
1810          * consumed or not. If reserves were used, a partial faulted mapping
1811          * at the time of fork() could consume its reserves on COW instead
1812          * of the full address range.
1813          */
1814         if (!(vma->vm_flags & VM_SHARED) &&
1815                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1816                         old_page != pagecache_page)
1817                 outside_reserve = 1;
1818
1819         page_cache_get(old_page);
1820         new_page = alloc_huge_page(vma, address, outside_reserve);
1821
1822         if (IS_ERR(new_page)) {
1823                 page_cache_release(old_page);
1824
1825                 /*
1826                  * If a process owning a MAP_PRIVATE mapping fails to COW,
1827                  * it is due to references held by a child and an insufficient
1828                  * huge page pool. To guarantee the original mappers
1829                  * reliability, unmap the page from child processes. The child
1830                  * may get SIGKILLed if it later faults.
1831                  */
1832                 if (outside_reserve) {
1833                         BUG_ON(huge_pte_none(pte));
1834                         if (unmap_ref_private(mm, vma, old_page, address)) {
1835                                 BUG_ON(page_count(old_page) != 1);
1836                                 BUG_ON(huge_pte_none(pte));
1837                                 goto retry_avoidcopy;
1838                         }
1839                         WARN_ON_ONCE(1);
1840                 }
1841
1842                 return -PTR_ERR(new_page);
1843         }
1844
1845         spin_unlock(&mm->page_table_lock);
1846         copy_huge_page(new_page, old_page, address, vma);
1847         __SetPageUptodate(new_page);
1848         spin_lock(&mm->page_table_lock);
1849
1850         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1851         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1852                 /* Break COW */
1853                 huge_ptep_clear_flush(vma, address, ptep);
1854                 set_huge_pte_at(mm, address, ptep,
1855                                 make_huge_pte(vma, new_page, 1));
1856                 /* Make the old page be freed below */
1857                 new_page = old_page;
1858         }
1859         page_cache_release(new_page);
1860         page_cache_release(old_page);
1861         return 0;
1862 }
1863
1864 /* Return the pagecache page at a given address within a VMA */
1865 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1866                         struct vm_area_struct *vma, unsigned long address)
1867 {
1868         struct address_space *mapping;
1869         pgoff_t idx;
1870
1871         mapping = vma->vm_file->f_mapping;
1872         idx = vma_hugecache_offset(h, vma, address);
1873
1874         return find_lock_page(mapping, idx);
1875 }
1876
1877 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1878                         unsigned long address, pte_t *ptep, int write_access)
1879 {
1880         struct hstate *h = hstate_vma(vma);
1881         int ret = VM_FAULT_SIGBUS;
1882         pgoff_t idx;
1883         unsigned long size;
1884         struct page *page;
1885         struct address_space *mapping;
1886         pte_t new_pte;
1887
1888         /*
1889          * Currently, we are forced to kill the process in the event the
1890          * original mapper has unmapped pages from the child due to a failed
1891          * COW. Warn that such a situation has occured as it may not be obvious
1892          */
1893         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1894                 printk(KERN_WARNING
1895                         "PID %d killed due to inadequate hugepage pool\n",
1896                         current->pid);
1897                 return ret;
1898         }
1899
1900         mapping = vma->vm_file->f_mapping;
1901         idx = vma_hugecache_offset(h, vma, address);
1902
1903         /*
1904          * Use page lock to guard against racing truncation
1905          * before we get page_table_lock.
1906          */
1907 retry:
1908         page = find_lock_page(mapping, idx);
1909         if (!page) {
1910                 size = i_size_read(mapping->host) >> huge_page_shift(h);
1911                 if (idx >= size)
1912                         goto out;
1913                 page = alloc_huge_page(vma, address, 0);
1914                 if (IS_ERR(page)) {
1915                         ret = -PTR_ERR(page);
1916                         goto out;
1917                 }
1918                 clear_huge_page(page, address, huge_page_size(h));
1919                 __SetPageUptodate(page);
1920
1921                 if (vma->vm_flags & VM_SHARED) {
1922                         int err;
1923                         struct inode *inode = mapping->host;
1924
1925                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1926                         if (err) {
1927                                 put_page(page);
1928                                 if (err == -EEXIST)
1929                                         goto retry;
1930                                 goto out;
1931                         }
1932
1933                         spin_lock(&inode->i_lock);
1934                         inode->i_blocks += blocks_per_huge_page(h);
1935                         spin_unlock(&inode->i_lock);
1936                 } else
1937                         lock_page(page);
1938         }
1939
1940         spin_lock(&mm->page_table_lock);
1941         size = i_size_read(mapping->host) >> huge_page_shift(h);
1942         if (idx >= size)
1943                 goto backout;
1944
1945         ret = 0;
1946         if (!huge_pte_none(huge_ptep_get(ptep)))
1947                 goto backout;
1948
1949         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
1950                                 && (vma->vm_flags & VM_SHARED)));
1951         set_huge_pte_at(mm, address, ptep, new_pte);
1952
1953         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1954                 /* Optimization, do the COW without a second fault */
1955                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
1956         }
1957
1958         spin_unlock(&mm->page_table_lock);
1959         unlock_page(page);
1960 out:
1961         return ret;
1962
1963 backout:
1964         spin_unlock(&mm->page_table_lock);
1965         unlock_page(page);
1966         put_page(page);
1967         goto out;
1968 }
1969
1970 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1971                         unsigned long address, int write_access)
1972 {
1973         pte_t *ptep;
1974         pte_t entry;
1975         int ret;
1976         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
1977         struct hstate *h = hstate_vma(vma);
1978
1979         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
1980         if (!ptep)
1981                 return VM_FAULT_OOM;
1982
1983         /*
1984          * Serialize hugepage allocation and instantiation, so that we don't
1985          * get spurious allocation failures if two CPUs race to instantiate
1986          * the same page in the page cache.
1987          */
1988         mutex_lock(&hugetlb_instantiation_mutex);
1989         entry = huge_ptep_get(ptep);
1990         if (huge_pte_none(entry)) {
1991                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1992                 mutex_unlock(&hugetlb_instantiation_mutex);
1993                 return ret;
1994         }
1995
1996         ret = 0;
1997
1998         spin_lock(&mm->page_table_lock);
1999         /* Check for a racing update before calling hugetlb_cow */
2000         if (likely(pte_same(entry, huge_ptep_get(ptep))))
2001                 if (write_access && !pte_write(entry)) {
2002                         struct page *page;
2003                         page = hugetlbfs_pagecache_page(h, vma, address);
2004                         ret = hugetlb_cow(mm, vma, address, ptep, entry, page);
2005                         if (page) {
2006                                 unlock_page(page);
2007                                 put_page(page);
2008                         }
2009                 }
2010         spin_unlock(&mm->page_table_lock);
2011         mutex_unlock(&hugetlb_instantiation_mutex);
2012
2013         return ret;
2014 }
2015
2016 /* Can be overriden by architectures */
2017 __attribute__((weak)) struct page *
2018 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2019                pud_t *pud, int write)
2020 {
2021         BUG();
2022         return NULL;
2023 }
2024
2025 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2026                         struct page **pages, struct vm_area_struct **vmas,
2027                         unsigned long *position, int *length, int i,
2028                         int write)
2029 {
2030         unsigned long pfn_offset;
2031         unsigned long vaddr = *position;
2032         int remainder = *length;
2033         struct hstate *h = hstate_vma(vma);
2034
2035         spin_lock(&mm->page_table_lock);
2036         while (vaddr < vma->vm_end && remainder) {
2037                 pte_t *pte;
2038                 struct page *page;
2039
2040                 /*
2041                  * Some archs (sparc64, sh*) have multiple pte_ts to
2042                  * each hugepage.  We have to make * sure we get the
2043                  * first, for the page indexing below to work.
2044                  */
2045                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2046
2047                 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
2048                     (write && !pte_write(huge_ptep_get(pte)))) {
2049                         int ret;
2050
2051                         spin_unlock(&mm->page_table_lock);
2052                         ret = hugetlb_fault(mm, vma, vaddr, write);
2053                         spin_lock(&mm->page_table_lock);
2054                         if (!(ret & VM_FAULT_ERROR))
2055                                 continue;
2056
2057                         remainder = 0;
2058                         if (!i)
2059                                 i = -EFAULT;
2060                         break;
2061                 }
2062
2063                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2064                 page = pte_page(huge_ptep_get(pte));
2065 same_page:
2066                 if (pages) {
2067                         get_page(page);
2068                         pages[i] = page + pfn_offset;
2069                 }
2070
2071                 if (vmas)
2072                         vmas[i] = vma;
2073
2074                 vaddr += PAGE_SIZE;
2075                 ++pfn_offset;
2076                 --remainder;
2077                 ++i;
2078                 if (vaddr < vma->vm_end && remainder &&
2079                                 pfn_offset < pages_per_huge_page(h)) {
2080                         /*
2081                          * We use pfn_offset to avoid touching the pageframes
2082                          * of this compound page.
2083                          */
2084                         goto same_page;
2085                 }
2086         }
2087         spin_unlock(&mm->page_table_lock);
2088         *length = remainder;
2089         *position = vaddr;
2090
2091         return i;
2092 }
2093
2094 void hugetlb_change_protection(struct vm_area_struct *vma,
2095                 unsigned long address, unsigned long end, pgprot_t newprot)
2096 {
2097         struct mm_struct *mm = vma->vm_mm;
2098         unsigned long start = address;
2099         pte_t *ptep;
2100         pte_t pte;
2101         struct hstate *h = hstate_vma(vma);
2102
2103         BUG_ON(address >= end);
2104         flush_cache_range(vma, address, end);
2105
2106         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2107         spin_lock(&mm->page_table_lock);
2108         for (; address < end; address += huge_page_size(h)) {
2109                 ptep = huge_pte_offset(mm, address);
2110                 if (!ptep)
2111                         continue;
2112                 if (huge_pmd_unshare(mm, &address, ptep))
2113                         continue;
2114                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2115                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2116                         pte = pte_mkhuge(pte_modify(pte, newprot));
2117                         set_huge_pte_at(mm, address, ptep, pte);
2118                 }
2119         }
2120         spin_unlock(&mm->page_table_lock);
2121         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2122
2123         flush_tlb_range(vma, start, end);
2124 }
2125
2126 int hugetlb_reserve_pages(struct inode *inode,
2127                                         long from, long to,
2128                                         struct vm_area_struct *vma)
2129 {
2130         long ret, chg;
2131         struct hstate *h = hstate_inode(inode);
2132
2133         if (vma && vma->vm_flags & VM_NORESERVE)
2134                 return 0;
2135
2136         /*
2137          * Shared mappings base their reservation on the number of pages that
2138          * are already allocated on behalf of the file. Private mappings need
2139          * to reserve the full area even if read-only as mprotect() may be
2140          * called to make the mapping read-write. Assume !vma is a shm mapping
2141          */
2142         if (!vma || vma->vm_flags & VM_SHARED)
2143                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2144         else {
2145                 struct resv_map *resv_map = resv_map_alloc();
2146                 if (!resv_map)
2147                         return -ENOMEM;
2148
2149                 chg = to - from;
2150
2151                 set_vma_resv_map(vma, resv_map);
2152                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2153         }
2154
2155         if (chg < 0)
2156                 return chg;
2157
2158         if (hugetlb_get_quota(inode->i_mapping, chg))
2159                 return -ENOSPC;
2160         ret = hugetlb_acct_memory(h, chg);
2161         if (ret < 0) {
2162                 hugetlb_put_quota(inode->i_mapping, chg);
2163                 return ret;
2164         }
2165         if (!vma || vma->vm_flags & VM_SHARED)
2166                 region_add(&inode->i_mapping->private_list, from, to);
2167         return 0;
2168 }
2169
2170 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2171 {
2172         struct hstate *h = hstate_inode(inode);
2173         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2174
2175         spin_lock(&inode->i_lock);
2176         inode->i_blocks -= blocks_per_huge_page(h);
2177         spin_unlock(&inode->i_lock);
2178
2179         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2180         hugetlb_acct_memory(h, -(chg - freed));
2181 }