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