proc: switch /proc/meminfo to seq_file
[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/seq_file.h>
11 #include <linux/sysctl.h>
12 #include <linux/highmem.h>
13 #include <linux/mmu_notifier.h>
14 #include <linux/nodemask.h>
15 #include <linux/pagemap.h>
16 #include <linux/mempolicy.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21
22 #include <asm/page.h>
23 #include <asm/pgtable.h>
24 #include <asm/io.h>
25
26 #include <linux/hugetlb.h>
27 #include "internal.h"
28
29 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
30 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
31 unsigned long hugepages_treat_as_movable;
32
33 static int max_hstate;
34 unsigned int default_hstate_idx;
35 struct hstate hstates[HUGE_MAX_HSTATE];
36
37 __initdata LIST_HEAD(huge_boot_pages);
38
39 /* for command line parsing */
40 static struct hstate * __initdata parsed_hstate;
41 static unsigned long __initdata default_hstate_max_huge_pages;
42 static unsigned long __initdata default_hstate_size;
43
44 #define for_each_hstate(h) \
45         for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
46
47 /*
48  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
49  */
50 static DEFINE_SPINLOCK(hugetlb_lock);
51
52 /*
53  * Region tracking -- allows tracking of reservations and instantiated pages
54  *                    across the pages in a mapping.
55  *
56  * The region data structures are protected by a combination of the mmap_sem
57  * and the hugetlb_instantion_mutex.  To access or modify a region the caller
58  * must either hold the mmap_sem for write, or the mmap_sem for read and
59  * the hugetlb_instantiation mutex:
60  *
61  *      down_write(&mm->mmap_sem);
62  * or
63  *      down_read(&mm->mmap_sem);
64  *      mutex_lock(&hugetlb_instantiation_mutex);
65  */
66 struct file_region {
67         struct list_head link;
68         long from;
69         long to;
70 };
71
72 static long region_add(struct list_head *head, long f, long t)
73 {
74         struct file_region *rg, *nrg, *trg;
75
76         /* Locate the region we are either in or before. */
77         list_for_each_entry(rg, head, link)
78                 if (f <= rg->to)
79                         break;
80
81         /* Round our left edge to the current segment if it encloses us. */
82         if (f > rg->from)
83                 f = rg->from;
84
85         /* Check for and consume any regions we now overlap with. */
86         nrg = rg;
87         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
88                 if (&rg->link == head)
89                         break;
90                 if (rg->from > t)
91                         break;
92
93                 /* If this area reaches higher then extend our area to
94                  * include it completely.  If this is not the first area
95                  * which we intend to reuse, free it. */
96                 if (rg->to > t)
97                         t = rg->to;
98                 if (rg != nrg) {
99                         list_del(&rg->link);
100                         kfree(rg);
101                 }
102         }
103         nrg->from = f;
104         nrg->to = t;
105         return 0;
106 }
107
108 static long region_chg(struct list_head *head, long f, long t)
109 {
110         struct file_region *rg, *nrg;
111         long chg = 0;
112
113         /* Locate the region we are before or in. */
114         list_for_each_entry(rg, head, link)
115                 if (f <= rg->to)
116                         break;
117
118         /* If we are below the current region then a new region is required.
119          * Subtle, allocate a new region at the position but make it zero
120          * size such that we can guarantee to record the reservation. */
121         if (&rg->link == head || t < rg->from) {
122                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
123                 if (!nrg)
124                         return -ENOMEM;
125                 nrg->from = f;
126                 nrg->to   = f;
127                 INIT_LIST_HEAD(&nrg->link);
128                 list_add(&nrg->link, rg->link.prev);
129
130                 return t - f;
131         }
132
133         /* Round our left edge to the current segment if it encloses us. */
134         if (f > rg->from)
135                 f = rg->from;
136         chg = t - f;
137
138         /* Check for and consume any regions we now overlap with. */
139         list_for_each_entry(rg, rg->link.prev, link) {
140                 if (&rg->link == head)
141                         break;
142                 if (rg->from > t)
143                         return chg;
144
145                 /* We overlap with this area, if it extends futher than
146                  * us then we must extend ourselves.  Account for its
147                  * existing reservation. */
148                 if (rg->to > t) {
149                         chg += rg->to - t;
150                         t = rg->to;
151                 }
152                 chg -= rg->to - rg->from;
153         }
154         return chg;
155 }
156
157 static long region_truncate(struct list_head *head, long end)
158 {
159         struct file_region *rg, *trg;
160         long chg = 0;
161
162         /* Locate the region we are either in or before. */
163         list_for_each_entry(rg, head, link)
164                 if (end <= rg->to)
165                         break;
166         if (&rg->link == head)
167                 return 0;
168
169         /* If we are in the middle of a region then adjust it. */
170         if (end > rg->from) {
171                 chg = rg->to - end;
172                 rg->to = end;
173                 rg = list_entry(rg->link.next, typeof(*rg), link);
174         }
175
176         /* Drop any remaining regions. */
177         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
178                 if (&rg->link == head)
179                         break;
180                 chg += rg->to - rg->from;
181                 list_del(&rg->link);
182                 kfree(rg);
183         }
184         return chg;
185 }
186
187 static long region_count(struct list_head *head, long f, long t)
188 {
189         struct file_region *rg;
190         long chg = 0;
191
192         /* Locate each segment we overlap with, and count that overlap. */
193         list_for_each_entry(rg, head, link) {
194                 int seg_from;
195                 int seg_to;
196
197                 if (rg->to <= f)
198                         continue;
199                 if (rg->from >= t)
200                         break;
201
202                 seg_from = max(rg->from, f);
203                 seg_to = min(rg->to, t);
204
205                 chg += seg_to - seg_from;
206         }
207
208         return chg;
209 }
210
211 /*
212  * Convert the address within this vma to the page offset within
213  * the mapping, in pagecache page units; huge pages here.
214  */
215 static pgoff_t vma_hugecache_offset(struct hstate *h,
216                         struct vm_area_struct *vma, unsigned long address)
217 {
218         return ((address - vma->vm_start) >> huge_page_shift(h)) +
219                         (vma->vm_pgoff >> huge_page_order(h));
220 }
221
222 /*
223  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
224  * bits of the reservation map pointer, which are always clear due to
225  * alignment.
226  */
227 #define HPAGE_RESV_OWNER    (1UL << 0)
228 #define HPAGE_RESV_UNMAPPED (1UL << 1)
229 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
230
231 /*
232  * These helpers are used to track how many pages are reserved for
233  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
234  * is guaranteed to have their future faults succeed.
235  *
236  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
237  * the reserve counters are updated with the hugetlb_lock held. It is safe
238  * to reset the VMA at fork() time as it is not in use yet and there is no
239  * chance of the global counters getting corrupted as a result of the values.
240  *
241  * The private mapping reservation is represented in a subtly different
242  * manner to a shared mapping.  A shared mapping has a region map associated
243  * with the underlying file, this region map represents the backing file
244  * pages which have ever had a reservation assigned which this persists even
245  * after the page is instantiated.  A private mapping has a region map
246  * associated with the original mmap which is attached to all VMAs which
247  * reference it, this region map represents those offsets which have consumed
248  * reservation ie. where pages have been instantiated.
249  */
250 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
251 {
252         return (unsigned long)vma->vm_private_data;
253 }
254
255 static void set_vma_private_data(struct vm_area_struct *vma,
256                                                         unsigned long value)
257 {
258         vma->vm_private_data = (void *)value;
259 }
260
261 struct resv_map {
262         struct kref refs;
263         struct list_head regions;
264 };
265
266 static struct resv_map *resv_map_alloc(void)
267 {
268         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
269         if (!resv_map)
270                 return NULL;
271
272         kref_init(&resv_map->refs);
273         INIT_LIST_HEAD(&resv_map->regions);
274
275         return resv_map;
276 }
277
278 static void resv_map_release(struct kref *ref)
279 {
280         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
281
282         /* Clear out any active regions before we release the map. */
283         region_truncate(&resv_map->regions, 0);
284         kfree(resv_map);
285 }
286
287 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
288 {
289         VM_BUG_ON(!is_vm_hugetlb_page(vma));
290         if (!(vma->vm_flags & VM_SHARED))
291                 return (struct resv_map *)(get_vma_private_data(vma) &
292                                                         ~HPAGE_RESV_MASK);
293         return NULL;
294 }
295
296 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
297 {
298         VM_BUG_ON(!is_vm_hugetlb_page(vma));
299         VM_BUG_ON(vma->vm_flags & VM_SHARED);
300
301         set_vma_private_data(vma, (get_vma_private_data(vma) &
302                                 HPAGE_RESV_MASK) | (unsigned long)map);
303 }
304
305 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
306 {
307         VM_BUG_ON(!is_vm_hugetlb_page(vma));
308         VM_BUG_ON(vma->vm_flags & VM_SHARED);
309
310         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
311 }
312
313 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
314 {
315         VM_BUG_ON(!is_vm_hugetlb_page(vma));
316
317         return (get_vma_private_data(vma) & flag) != 0;
318 }
319
320 /* Decrement the reserved pages in the hugepage pool by one */
321 static void decrement_hugepage_resv_vma(struct hstate *h,
322                         struct vm_area_struct *vma)
323 {
324         if (vma->vm_flags & VM_NORESERVE)
325                 return;
326
327         if (vma->vm_flags & VM_SHARED) {
328                 /* Shared mappings always use reserves */
329                 h->resv_huge_pages--;
330         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
331                 /*
332                  * Only the process that called mmap() has reserves for
333                  * private mappings.
334                  */
335                 h->resv_huge_pages--;
336         }
337 }
338
339 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
340 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
341 {
342         VM_BUG_ON(!is_vm_hugetlb_page(vma));
343         if (!(vma->vm_flags & VM_SHARED))
344                 vma->vm_private_data = (void *)0;
345 }
346
347 /* Returns true if the VMA has associated reserve pages */
348 static int vma_has_reserves(struct vm_area_struct *vma)
349 {
350         if (vma->vm_flags & VM_SHARED)
351                 return 1;
352         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
353                 return 1;
354         return 0;
355 }
356
357 static void clear_huge_page(struct page *page,
358                         unsigned long addr, unsigned long sz)
359 {
360         int i;
361
362         might_sleep();
363         for (i = 0; i < sz/PAGE_SIZE; i++) {
364                 cond_resched();
365                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
366         }
367 }
368
369 static void copy_huge_page(struct page *dst, struct page *src,
370                            unsigned long addr, struct vm_area_struct *vma)
371 {
372         int i;
373         struct hstate *h = hstate_vma(vma);
374
375         might_sleep();
376         for (i = 0; i < pages_per_huge_page(h); i++) {
377                 cond_resched();
378                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
379         }
380 }
381
382 static void enqueue_huge_page(struct hstate *h, struct page *page)
383 {
384         int nid = page_to_nid(page);
385         list_add(&page->lru, &h->hugepage_freelists[nid]);
386         h->free_huge_pages++;
387         h->free_huge_pages_node[nid]++;
388 }
389
390 static struct page *dequeue_huge_page(struct hstate *h)
391 {
392         int nid;
393         struct page *page = NULL;
394
395         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
396                 if (!list_empty(&h->hugepage_freelists[nid])) {
397                         page = list_entry(h->hugepage_freelists[nid].next,
398                                           struct page, lru);
399                         list_del(&page->lru);
400                         h->free_huge_pages--;
401                         h->free_huge_pages_node[nid]--;
402                         break;
403                 }
404         }
405         return page;
406 }
407
408 static struct page *dequeue_huge_page_vma(struct hstate *h,
409                                 struct vm_area_struct *vma,
410                                 unsigned long address, int avoid_reserve)
411 {
412         int nid;
413         struct page *page = NULL;
414         struct mempolicy *mpol;
415         nodemask_t *nodemask;
416         struct zonelist *zonelist = huge_zonelist(vma, address,
417                                         htlb_alloc_mask, &mpol, &nodemask);
418         struct zone *zone;
419         struct zoneref *z;
420
421         /*
422          * A child process with MAP_PRIVATE mappings created by their parent
423          * have no page reserves. This check ensures that reservations are
424          * not "stolen". The child may still get SIGKILLed
425          */
426         if (!vma_has_reserves(vma) &&
427                         h->free_huge_pages - h->resv_huge_pages == 0)
428                 return NULL;
429
430         /* If reserves cannot be used, ensure enough pages are in the pool */
431         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
432                 return NULL;
433
434         for_each_zone_zonelist_nodemask(zone, z, zonelist,
435                                                 MAX_NR_ZONES - 1, nodemask) {
436                 nid = zone_to_nid(zone);
437                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
438                     !list_empty(&h->hugepage_freelists[nid])) {
439                         page = list_entry(h->hugepage_freelists[nid].next,
440                                           struct page, lru);
441                         list_del(&page->lru);
442                         h->free_huge_pages--;
443                         h->free_huge_pages_node[nid]--;
444
445                         if (!avoid_reserve)
446                                 decrement_hugepage_resv_vma(h, vma);
447
448                         break;
449                 }
450         }
451         mpol_cond_put(mpol);
452         return page;
453 }
454
455 static void update_and_free_page(struct hstate *h, struct page *page)
456 {
457         int i;
458
459         h->nr_huge_pages--;
460         h->nr_huge_pages_node[page_to_nid(page)]--;
461         for (i = 0; i < pages_per_huge_page(h); i++) {
462                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
463                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
464                                 1 << PG_private | 1<< PG_writeback);
465         }
466         set_compound_page_dtor(page, NULL);
467         set_page_refcounted(page);
468         arch_release_hugepage(page);
469         __free_pages(page, huge_page_order(h));
470 }
471
472 struct hstate *size_to_hstate(unsigned long size)
473 {
474         struct hstate *h;
475
476         for_each_hstate(h) {
477                 if (huge_page_size(h) == size)
478                         return h;
479         }
480         return NULL;
481 }
482
483 static void free_huge_page(struct page *page)
484 {
485         /*
486          * Can't pass hstate in here because it is called from the
487          * compound page destructor.
488          */
489         struct hstate *h = page_hstate(page);
490         int nid = page_to_nid(page);
491         struct address_space *mapping;
492
493         mapping = (struct address_space *) page_private(page);
494         set_page_private(page, 0);
495         BUG_ON(page_count(page));
496         INIT_LIST_HEAD(&page->lru);
497
498         spin_lock(&hugetlb_lock);
499         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
500                 update_and_free_page(h, page);
501                 h->surplus_huge_pages--;
502                 h->surplus_huge_pages_node[nid]--;
503         } else {
504                 enqueue_huge_page(h, page);
505         }
506         spin_unlock(&hugetlb_lock);
507         if (mapping)
508                 hugetlb_put_quota(mapping, 1);
509 }
510
511 /*
512  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
513  * balanced by operating on them in a round-robin fashion.
514  * Returns 1 if an adjustment was made.
515  */
516 static int adjust_pool_surplus(struct hstate *h, int delta)
517 {
518         static int prev_nid;
519         int nid = prev_nid;
520         int ret = 0;
521
522         VM_BUG_ON(delta != -1 && delta != 1);
523         do {
524                 nid = next_node(nid, node_online_map);
525                 if (nid == MAX_NUMNODES)
526                         nid = first_node(node_online_map);
527
528                 /* To shrink on this node, there must be a surplus page */
529                 if (delta < 0 && !h->surplus_huge_pages_node[nid])
530                         continue;
531                 /* Surplus cannot exceed the total number of pages */
532                 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
533                                                 h->nr_huge_pages_node[nid])
534                         continue;
535
536                 h->surplus_huge_pages += delta;
537                 h->surplus_huge_pages_node[nid] += delta;
538                 ret = 1;
539                 break;
540         } while (nid != prev_nid);
541
542         prev_nid = nid;
543         return ret;
544 }
545
546 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
547 {
548         set_compound_page_dtor(page, free_huge_page);
549         spin_lock(&hugetlb_lock);
550         h->nr_huge_pages++;
551         h->nr_huge_pages_node[nid]++;
552         spin_unlock(&hugetlb_lock);
553         put_page(page); /* free it into the hugepage allocator */
554 }
555
556 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
557 {
558         struct page *page;
559
560         if (h->order >= MAX_ORDER)
561                 return NULL;
562
563         page = alloc_pages_node(nid,
564                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
565                                                 __GFP_REPEAT|__GFP_NOWARN,
566                 huge_page_order(h));
567         if (page) {
568                 if (arch_prepare_hugepage(page)) {
569                         __free_pages(page, huge_page_order(h));
570                         return NULL;
571                 }
572                 prep_new_huge_page(h, page, nid);
573         }
574
575         return page;
576 }
577
578 /*
579  * Use a helper variable to find the next node and then
580  * copy it back to hugetlb_next_nid afterwards:
581  * otherwise there's a window in which a racer might
582  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
583  * But we don't need to use a spin_lock here: it really
584  * doesn't matter if occasionally a racer chooses the
585  * same nid as we do.  Move nid forward in the mask even
586  * if we just successfully allocated a hugepage so that
587  * the next caller gets hugepages on the next node.
588  */
589 static int hstate_next_node(struct hstate *h)
590 {
591         int next_nid;
592         next_nid = next_node(h->hugetlb_next_nid, node_online_map);
593         if (next_nid == MAX_NUMNODES)
594                 next_nid = first_node(node_online_map);
595         h->hugetlb_next_nid = next_nid;
596         return next_nid;
597 }
598
599 static int alloc_fresh_huge_page(struct hstate *h)
600 {
601         struct page *page;
602         int start_nid;
603         int next_nid;
604         int ret = 0;
605
606         start_nid = h->hugetlb_next_nid;
607
608         do {
609                 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
610                 if (page)
611                         ret = 1;
612                 next_nid = hstate_next_node(h);
613         } while (!page && h->hugetlb_next_nid != start_nid);
614
615         if (ret)
616                 count_vm_event(HTLB_BUDDY_PGALLOC);
617         else
618                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
619
620         return ret;
621 }
622
623 static struct page *alloc_buddy_huge_page(struct hstate *h,
624                         struct vm_area_struct *vma, unsigned long address)
625 {
626         struct page *page;
627         unsigned int nid;
628
629         if (h->order >= MAX_ORDER)
630                 return NULL;
631
632         /*
633          * Assume we will successfully allocate the surplus page to
634          * prevent racing processes from causing the surplus to exceed
635          * overcommit
636          *
637          * This however introduces a different race, where a process B
638          * tries to grow the static hugepage pool while alloc_pages() is
639          * called by process A. B will only examine the per-node
640          * counters in determining if surplus huge pages can be
641          * converted to normal huge pages in adjust_pool_surplus(). A
642          * won't be able to increment the per-node counter, until the
643          * lock is dropped by B, but B doesn't drop hugetlb_lock until
644          * no more huge pages can be converted from surplus to normal
645          * state (and doesn't try to convert again). Thus, we have a
646          * case where a surplus huge page exists, the pool is grown, and
647          * the surplus huge page still exists after, even though it
648          * should just have been converted to a normal huge page. This
649          * does not leak memory, though, as the hugepage will be freed
650          * once it is out of use. It also does not allow the counters to
651          * go out of whack in adjust_pool_surplus() as we don't modify
652          * the node values until we've gotten the hugepage and only the
653          * per-node value is checked there.
654          */
655         spin_lock(&hugetlb_lock);
656         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
657                 spin_unlock(&hugetlb_lock);
658                 return NULL;
659         } else {
660                 h->nr_huge_pages++;
661                 h->surplus_huge_pages++;
662         }
663         spin_unlock(&hugetlb_lock);
664
665         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
666                                         __GFP_REPEAT|__GFP_NOWARN,
667                                         huge_page_order(h));
668
669         if (page && arch_prepare_hugepage(page)) {
670                 __free_pages(page, huge_page_order(h));
671                 return NULL;
672         }
673
674         spin_lock(&hugetlb_lock);
675         if (page) {
676                 /*
677                  * This page is now managed by the hugetlb allocator and has
678                  * no users -- drop the buddy allocator's reference.
679                  */
680                 put_page_testzero(page);
681                 VM_BUG_ON(page_count(page));
682                 nid = page_to_nid(page);
683                 set_compound_page_dtor(page, free_huge_page);
684                 /*
685                  * We incremented the global counters already
686                  */
687                 h->nr_huge_pages_node[nid]++;
688                 h->surplus_huge_pages_node[nid]++;
689                 __count_vm_event(HTLB_BUDDY_PGALLOC);
690         } else {
691                 h->nr_huge_pages--;
692                 h->surplus_huge_pages--;
693                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
694         }
695         spin_unlock(&hugetlb_lock);
696
697         return page;
698 }
699
700 /*
701  * Increase the hugetlb pool such that it can accomodate a reservation
702  * of size 'delta'.
703  */
704 static int gather_surplus_pages(struct hstate *h, int delta)
705 {
706         struct list_head surplus_list;
707         struct page *page, *tmp;
708         int ret, i;
709         int needed, allocated;
710
711         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
712         if (needed <= 0) {
713                 h->resv_huge_pages += delta;
714                 return 0;
715         }
716
717         allocated = 0;
718         INIT_LIST_HEAD(&surplus_list);
719
720         ret = -ENOMEM;
721 retry:
722         spin_unlock(&hugetlb_lock);
723         for (i = 0; i < needed; i++) {
724                 page = alloc_buddy_huge_page(h, NULL, 0);
725                 if (!page) {
726                         /*
727                          * We were not able to allocate enough pages to
728                          * satisfy the entire reservation so we free what
729                          * we've allocated so far.
730                          */
731                         spin_lock(&hugetlb_lock);
732                         needed = 0;
733                         goto free;
734                 }
735
736                 list_add(&page->lru, &surplus_list);
737         }
738         allocated += needed;
739
740         /*
741          * After retaking hugetlb_lock, we need to recalculate 'needed'
742          * because either resv_huge_pages or free_huge_pages may have changed.
743          */
744         spin_lock(&hugetlb_lock);
745         needed = (h->resv_huge_pages + delta) -
746                         (h->free_huge_pages + allocated);
747         if (needed > 0)
748                 goto retry;
749
750         /*
751          * The surplus_list now contains _at_least_ the number of extra pages
752          * needed to accomodate the reservation.  Add the appropriate number
753          * of pages to the hugetlb pool and free the extras back to the buddy
754          * allocator.  Commit the entire reservation here to prevent another
755          * process from stealing the pages as they are added to the pool but
756          * before they are reserved.
757          */
758         needed += allocated;
759         h->resv_huge_pages += delta;
760         ret = 0;
761 free:
762         /* Free the needed pages to the hugetlb pool */
763         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
764                 if ((--needed) < 0)
765                         break;
766                 list_del(&page->lru);
767                 enqueue_huge_page(h, page);
768         }
769
770         /* Free unnecessary surplus pages to the buddy allocator */
771         if (!list_empty(&surplus_list)) {
772                 spin_unlock(&hugetlb_lock);
773                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
774                         list_del(&page->lru);
775                         /*
776                          * The page has a reference count of zero already, so
777                          * call free_huge_page directly instead of using
778                          * put_page.  This must be done with hugetlb_lock
779                          * unlocked which is safe because free_huge_page takes
780                          * hugetlb_lock before deciding how to free the page.
781                          */
782                         free_huge_page(page);
783                 }
784                 spin_lock(&hugetlb_lock);
785         }
786
787         return ret;
788 }
789
790 /*
791  * When releasing a hugetlb pool reservation, any surplus pages that were
792  * allocated to satisfy the reservation must be explicitly freed if they were
793  * never used.
794  */
795 static void return_unused_surplus_pages(struct hstate *h,
796                                         unsigned long unused_resv_pages)
797 {
798         static int nid = -1;
799         struct page *page;
800         unsigned long nr_pages;
801
802         /*
803          * We want to release as many surplus pages as possible, spread
804          * evenly across all nodes. Iterate across all nodes until we
805          * can no longer free unreserved surplus pages. This occurs when
806          * the nodes with surplus pages have no free pages.
807          */
808         unsigned long remaining_iterations = num_online_nodes();
809
810         /* Uncommit the reservation */
811         h->resv_huge_pages -= unused_resv_pages;
812
813         /* Cannot return gigantic pages currently */
814         if (h->order >= MAX_ORDER)
815                 return;
816
817         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
818
819         while (remaining_iterations-- && nr_pages) {
820                 nid = next_node(nid, node_online_map);
821                 if (nid == MAX_NUMNODES)
822                         nid = first_node(node_online_map);
823
824                 if (!h->surplus_huge_pages_node[nid])
825                         continue;
826
827                 if (!list_empty(&h->hugepage_freelists[nid])) {
828                         page = list_entry(h->hugepage_freelists[nid].next,
829                                           struct page, lru);
830                         list_del(&page->lru);
831                         update_and_free_page(h, page);
832                         h->free_huge_pages--;
833                         h->free_huge_pages_node[nid]--;
834                         h->surplus_huge_pages--;
835                         h->surplus_huge_pages_node[nid]--;
836                         nr_pages--;
837                         remaining_iterations = num_online_nodes();
838                 }
839         }
840 }
841
842 /*
843  * Determine if the huge page at addr within the vma has an associated
844  * reservation.  Where it does not we will need to logically increase
845  * reservation and actually increase quota before an allocation can occur.
846  * Where any new reservation would be required the reservation change is
847  * prepared, but not committed.  Once the page has been quota'd allocated
848  * an instantiated the change should be committed via vma_commit_reservation.
849  * No action is required on failure.
850  */
851 static int vma_needs_reservation(struct hstate *h,
852                         struct vm_area_struct *vma, unsigned long addr)
853 {
854         struct address_space *mapping = vma->vm_file->f_mapping;
855         struct inode *inode = mapping->host;
856
857         if (vma->vm_flags & VM_SHARED) {
858                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
859                 return region_chg(&inode->i_mapping->private_list,
860                                                         idx, idx + 1);
861
862         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
863                 return 1;
864
865         } else  {
866                 int err;
867                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
868                 struct resv_map *reservations = vma_resv_map(vma);
869
870                 err = region_chg(&reservations->regions, idx, idx + 1);
871                 if (err < 0)
872                         return err;
873                 return 0;
874         }
875 }
876 static void vma_commit_reservation(struct hstate *h,
877                         struct vm_area_struct *vma, unsigned long addr)
878 {
879         struct address_space *mapping = vma->vm_file->f_mapping;
880         struct inode *inode = mapping->host;
881
882         if (vma->vm_flags & VM_SHARED) {
883                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
884                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
885
886         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
887                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
888                 struct resv_map *reservations = vma_resv_map(vma);
889
890                 /* Mark this page used in the map. */
891                 region_add(&reservations->regions, idx, idx + 1);
892         }
893 }
894
895 static struct page *alloc_huge_page(struct vm_area_struct *vma,
896                                     unsigned long addr, int avoid_reserve)
897 {
898         struct hstate *h = hstate_vma(vma);
899         struct page *page;
900         struct address_space *mapping = vma->vm_file->f_mapping;
901         struct inode *inode = mapping->host;
902         unsigned int chg;
903
904         /*
905          * Processes that did not create the mapping will have no reserves and
906          * will not have accounted against quota. Check that the quota can be
907          * made before satisfying the allocation
908          * MAP_NORESERVE mappings may also need pages and quota allocated
909          * if no reserve mapping overlaps.
910          */
911         chg = vma_needs_reservation(h, vma, addr);
912         if (chg < 0)
913                 return ERR_PTR(chg);
914         if (chg)
915                 if (hugetlb_get_quota(inode->i_mapping, chg))
916                         return ERR_PTR(-ENOSPC);
917
918         spin_lock(&hugetlb_lock);
919         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
920         spin_unlock(&hugetlb_lock);
921
922         if (!page) {
923                 page = alloc_buddy_huge_page(h, vma, addr);
924                 if (!page) {
925                         hugetlb_put_quota(inode->i_mapping, chg);
926                         return ERR_PTR(-VM_FAULT_OOM);
927                 }
928         }
929
930         set_page_refcounted(page);
931         set_page_private(page, (unsigned long) mapping);
932
933         vma_commit_reservation(h, vma, addr);
934
935         return page;
936 }
937
938 __attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
939 {
940         struct huge_bootmem_page *m;
941         int nr_nodes = nodes_weight(node_online_map);
942
943         while (nr_nodes) {
944                 void *addr;
945
946                 addr = __alloc_bootmem_node_nopanic(
947                                 NODE_DATA(h->hugetlb_next_nid),
948                                 huge_page_size(h), huge_page_size(h), 0);
949
950                 if (addr) {
951                         /*
952                          * Use the beginning of the huge page to store the
953                          * huge_bootmem_page struct (until gather_bootmem
954                          * puts them into the mem_map).
955                          */
956                         m = addr;
957                         if (m)
958                                 goto found;
959                 }
960                 hstate_next_node(h);
961                 nr_nodes--;
962         }
963         return 0;
964
965 found:
966         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
967         /* Put them into a private list first because mem_map is not up yet */
968         list_add(&m->list, &huge_boot_pages);
969         m->hstate = h;
970         return 1;
971 }
972
973 /* Put bootmem huge pages into the standard lists after mem_map is up */
974 static void __init gather_bootmem_prealloc(void)
975 {
976         struct huge_bootmem_page *m;
977
978         list_for_each_entry(m, &huge_boot_pages, list) {
979                 struct page *page = virt_to_page(m);
980                 struct hstate *h = m->hstate;
981                 __ClearPageReserved(page);
982                 WARN_ON(page_count(page) != 1);
983                 prep_compound_page(page, h->order);
984                 prep_new_huge_page(h, page, page_to_nid(page));
985         }
986 }
987
988 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
989 {
990         unsigned long i;
991
992         for (i = 0; i < h->max_huge_pages; ++i) {
993                 if (h->order >= MAX_ORDER) {
994                         if (!alloc_bootmem_huge_page(h))
995                                 break;
996                 } else if (!alloc_fresh_huge_page(h))
997                         break;
998         }
999         h->max_huge_pages = i;
1000 }
1001
1002 static void __init hugetlb_init_hstates(void)
1003 {
1004         struct hstate *h;
1005
1006         for_each_hstate(h) {
1007                 /* oversize hugepages were init'ed in early boot */
1008                 if (h->order < MAX_ORDER)
1009                         hugetlb_hstate_alloc_pages(h);
1010         }
1011 }
1012
1013 static char * __init memfmt(char *buf, unsigned long n)
1014 {
1015         if (n >= (1UL << 30))
1016                 sprintf(buf, "%lu GB", n >> 30);
1017         else if (n >= (1UL << 20))
1018                 sprintf(buf, "%lu MB", n >> 20);
1019         else
1020                 sprintf(buf, "%lu KB", n >> 10);
1021         return buf;
1022 }
1023
1024 static void __init report_hugepages(void)
1025 {
1026         struct hstate *h;
1027
1028         for_each_hstate(h) {
1029                 char buf[32];
1030                 printk(KERN_INFO "HugeTLB registered %s page size, "
1031                                  "pre-allocated %ld pages\n",
1032                         memfmt(buf, huge_page_size(h)),
1033                         h->free_huge_pages);
1034         }
1035 }
1036
1037 #ifdef CONFIG_HIGHMEM
1038 static void try_to_free_low(struct hstate *h, unsigned long count)
1039 {
1040         int i;
1041
1042         if (h->order >= MAX_ORDER)
1043                 return;
1044
1045         for (i = 0; i < MAX_NUMNODES; ++i) {
1046                 struct page *page, *next;
1047                 struct list_head *freel = &h->hugepage_freelists[i];
1048                 list_for_each_entry_safe(page, next, freel, lru) {
1049                         if (count >= h->nr_huge_pages)
1050                                 return;
1051                         if (PageHighMem(page))
1052                                 continue;
1053                         list_del(&page->lru);
1054                         update_and_free_page(h, page);
1055                         h->free_huge_pages--;
1056                         h->free_huge_pages_node[page_to_nid(page)]--;
1057                 }
1058         }
1059 }
1060 #else
1061 static inline void try_to_free_low(struct hstate *h, unsigned long count)
1062 {
1063 }
1064 #endif
1065
1066 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1067 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
1068 {
1069         unsigned long min_count, ret;
1070
1071         if (h->order >= MAX_ORDER)
1072                 return h->max_huge_pages;
1073
1074         /*
1075          * Increase the pool size
1076          * First take pages out of surplus state.  Then make up the
1077          * remaining difference by allocating fresh huge pages.
1078          *
1079          * We might race with alloc_buddy_huge_page() here and be unable
1080          * to convert a surplus huge page to a normal huge page. That is
1081          * not critical, though, it just means the overall size of the
1082          * pool might be one hugepage larger than it needs to be, but
1083          * within all the constraints specified by the sysctls.
1084          */
1085         spin_lock(&hugetlb_lock);
1086         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1087                 if (!adjust_pool_surplus(h, -1))
1088                         break;
1089         }
1090
1091         while (count > persistent_huge_pages(h)) {
1092                 /*
1093                  * If this allocation races such that we no longer need the
1094                  * page, free_huge_page will handle it by freeing the page
1095                  * and reducing the surplus.
1096                  */
1097                 spin_unlock(&hugetlb_lock);
1098                 ret = alloc_fresh_huge_page(h);
1099                 spin_lock(&hugetlb_lock);
1100                 if (!ret)
1101                         goto out;
1102
1103         }
1104
1105         /*
1106          * Decrease the pool size
1107          * First return free pages to the buddy allocator (being careful
1108          * to keep enough around to satisfy reservations).  Then place
1109          * pages into surplus state as needed so the pool will shrink
1110          * to the desired size as pages become free.
1111          *
1112          * By placing pages into the surplus state independent of the
1113          * overcommit value, we are allowing the surplus pool size to
1114          * exceed overcommit. There are few sane options here. Since
1115          * alloc_buddy_huge_page() is checking the global counter,
1116          * though, we'll note that we're not allowed to exceed surplus
1117          * and won't grow the pool anywhere else. Not until one of the
1118          * sysctls are changed, or the surplus pages go out of use.
1119          */
1120         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1121         min_count = max(count, min_count);
1122         try_to_free_low(h, min_count);
1123         while (min_count < persistent_huge_pages(h)) {
1124                 struct page *page = dequeue_huge_page(h);
1125                 if (!page)
1126                         break;
1127                 update_and_free_page(h, page);
1128         }
1129         while (count < persistent_huge_pages(h)) {
1130                 if (!adjust_pool_surplus(h, 1))
1131                         break;
1132         }
1133 out:
1134         ret = persistent_huge_pages(h);
1135         spin_unlock(&hugetlb_lock);
1136         return ret;
1137 }
1138
1139 #define HSTATE_ATTR_RO(_name) \
1140         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1141
1142 #define HSTATE_ATTR(_name) \
1143         static struct kobj_attribute _name##_attr = \
1144                 __ATTR(_name, 0644, _name##_show, _name##_store)
1145
1146 static struct kobject *hugepages_kobj;
1147 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1148
1149 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1150 {
1151         int i;
1152         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1153                 if (hstate_kobjs[i] == kobj)
1154                         return &hstates[i];
1155         BUG();
1156         return NULL;
1157 }
1158
1159 static ssize_t nr_hugepages_show(struct kobject *kobj,
1160                                         struct kobj_attribute *attr, char *buf)
1161 {
1162         struct hstate *h = kobj_to_hstate(kobj);
1163         return sprintf(buf, "%lu\n", h->nr_huge_pages);
1164 }
1165 static ssize_t nr_hugepages_store(struct kobject *kobj,
1166                 struct kobj_attribute *attr, const char *buf, size_t count)
1167 {
1168         int err;
1169         unsigned long input;
1170         struct hstate *h = kobj_to_hstate(kobj);
1171
1172         err = strict_strtoul(buf, 10, &input);
1173         if (err)
1174                 return 0;
1175
1176         h->max_huge_pages = set_max_huge_pages(h, input);
1177
1178         return count;
1179 }
1180 HSTATE_ATTR(nr_hugepages);
1181
1182 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1183                                         struct kobj_attribute *attr, char *buf)
1184 {
1185         struct hstate *h = kobj_to_hstate(kobj);
1186         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1187 }
1188 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1189                 struct kobj_attribute *attr, const char *buf, size_t count)
1190 {
1191         int err;
1192         unsigned long input;
1193         struct hstate *h = kobj_to_hstate(kobj);
1194
1195         err = strict_strtoul(buf, 10, &input);
1196         if (err)
1197                 return 0;
1198
1199         spin_lock(&hugetlb_lock);
1200         h->nr_overcommit_huge_pages = input;
1201         spin_unlock(&hugetlb_lock);
1202
1203         return count;
1204 }
1205 HSTATE_ATTR(nr_overcommit_hugepages);
1206
1207 static ssize_t free_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->free_huge_pages);
1212 }
1213 HSTATE_ATTR_RO(free_hugepages);
1214
1215 static ssize_t resv_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->resv_huge_pages);
1220 }
1221 HSTATE_ATTR_RO(resv_hugepages);
1222
1223 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1224                                         struct kobj_attribute *attr, char *buf)
1225 {
1226         struct hstate *h = kobj_to_hstate(kobj);
1227         return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1228 }
1229 HSTATE_ATTR_RO(surplus_hugepages);
1230
1231 static struct attribute *hstate_attrs[] = {
1232         &nr_hugepages_attr.attr,
1233         &nr_overcommit_hugepages_attr.attr,
1234         &free_hugepages_attr.attr,
1235         &resv_hugepages_attr.attr,
1236         &surplus_hugepages_attr.attr,
1237         NULL,
1238 };
1239
1240 static struct attribute_group hstate_attr_group = {
1241         .attrs = hstate_attrs,
1242 };
1243
1244 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1245 {
1246         int retval;
1247
1248         hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1249                                                         hugepages_kobj);
1250         if (!hstate_kobjs[h - hstates])
1251                 return -ENOMEM;
1252
1253         retval = sysfs_create_group(hstate_kobjs[h - hstates],
1254                                                         &hstate_attr_group);
1255         if (retval)
1256                 kobject_put(hstate_kobjs[h - hstates]);
1257
1258         return retval;
1259 }
1260
1261 static void __init hugetlb_sysfs_init(void)
1262 {
1263         struct hstate *h;
1264         int err;
1265
1266         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1267         if (!hugepages_kobj)
1268                 return;
1269
1270         for_each_hstate(h) {
1271                 err = hugetlb_sysfs_add_hstate(h);
1272                 if (err)
1273                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1274                                                                 h->name);
1275         }
1276 }
1277
1278 static void __exit hugetlb_exit(void)
1279 {
1280         struct hstate *h;
1281
1282         for_each_hstate(h) {
1283                 kobject_put(hstate_kobjs[h - hstates]);
1284         }
1285
1286         kobject_put(hugepages_kobj);
1287 }
1288 module_exit(hugetlb_exit);
1289
1290 static int __init hugetlb_init(void)
1291 {
1292         /* Some platform decide whether they support huge pages at boot
1293          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1294          * there is no such support
1295          */
1296         if (HPAGE_SHIFT == 0)
1297                 return 0;
1298
1299         if (!size_to_hstate(default_hstate_size)) {
1300                 default_hstate_size = HPAGE_SIZE;
1301                 if (!size_to_hstate(default_hstate_size))
1302                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1303         }
1304         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1305         if (default_hstate_max_huge_pages)
1306                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1307
1308         hugetlb_init_hstates();
1309
1310         gather_bootmem_prealloc();
1311
1312         report_hugepages();
1313
1314         hugetlb_sysfs_init();
1315
1316         return 0;
1317 }
1318 module_init(hugetlb_init);
1319
1320 /* Should be called on processing a hugepagesz=... option */
1321 void __init hugetlb_add_hstate(unsigned order)
1322 {
1323         struct hstate *h;
1324         unsigned long i;
1325
1326         if (size_to_hstate(PAGE_SIZE << order)) {
1327                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1328                 return;
1329         }
1330         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1331         BUG_ON(order == 0);
1332         h = &hstates[max_hstate++];
1333         h->order = order;
1334         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1335         h->nr_huge_pages = 0;
1336         h->free_huge_pages = 0;
1337         for (i = 0; i < MAX_NUMNODES; ++i)
1338                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1339         h->hugetlb_next_nid = first_node(node_online_map);
1340         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1341                                         huge_page_size(h)/1024);
1342
1343         parsed_hstate = h;
1344 }
1345
1346 static int __init hugetlb_nrpages_setup(char *s)
1347 {
1348         unsigned long *mhp;
1349         static unsigned long *last_mhp;
1350
1351         /*
1352          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1353          * so this hugepages= parameter goes to the "default hstate".
1354          */
1355         if (!max_hstate)
1356                 mhp = &default_hstate_max_huge_pages;
1357         else
1358                 mhp = &parsed_hstate->max_huge_pages;
1359
1360         if (mhp == last_mhp) {
1361                 printk(KERN_WARNING "hugepages= specified twice without "
1362                         "interleaving hugepagesz=, ignoring\n");
1363                 return 1;
1364         }
1365
1366         if (sscanf(s, "%lu", mhp) <= 0)
1367                 *mhp = 0;
1368
1369         /*
1370          * Global state is always initialized later in hugetlb_init.
1371          * But we need to allocate >= MAX_ORDER hstates here early to still
1372          * use the bootmem allocator.
1373          */
1374         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1375                 hugetlb_hstate_alloc_pages(parsed_hstate);
1376
1377         last_mhp = mhp;
1378
1379         return 1;
1380 }
1381 __setup("hugepages=", hugetlb_nrpages_setup);
1382
1383 static int __init hugetlb_default_setup(char *s)
1384 {
1385         default_hstate_size = memparse(s, &s);
1386         return 1;
1387 }
1388 __setup("default_hugepagesz=", hugetlb_default_setup);
1389
1390 static unsigned int cpuset_mems_nr(unsigned int *array)
1391 {
1392         int node;
1393         unsigned int nr = 0;
1394
1395         for_each_node_mask(node, cpuset_current_mems_allowed)
1396                 nr += array[node];
1397
1398         return nr;
1399 }
1400
1401 #ifdef CONFIG_SYSCTL
1402 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1403                            struct file *file, void __user *buffer,
1404                            size_t *length, loff_t *ppos)
1405 {
1406         struct hstate *h = &default_hstate;
1407         unsigned long tmp;
1408
1409         if (!write)
1410                 tmp = h->max_huge_pages;
1411
1412         table->data = &tmp;
1413         table->maxlen = sizeof(unsigned long);
1414         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1415
1416         if (write)
1417                 h->max_huge_pages = set_max_huge_pages(h, tmp);
1418
1419         return 0;
1420 }
1421
1422 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1423                         struct file *file, void __user *buffer,
1424                         size_t *length, loff_t *ppos)
1425 {
1426         proc_dointvec(table, write, file, buffer, length, ppos);
1427         if (hugepages_treat_as_movable)
1428                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1429         else
1430                 htlb_alloc_mask = GFP_HIGHUSER;
1431         return 0;
1432 }
1433
1434 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1435                         struct file *file, void __user *buffer,
1436                         size_t *length, loff_t *ppos)
1437 {
1438         struct hstate *h = &default_hstate;
1439         unsigned long tmp;
1440
1441         if (!write)
1442                 tmp = h->nr_overcommit_huge_pages;
1443
1444         table->data = &tmp;
1445         table->maxlen = sizeof(unsigned long);
1446         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1447
1448         if (write) {
1449                 spin_lock(&hugetlb_lock);
1450                 h->nr_overcommit_huge_pages = tmp;
1451                 spin_unlock(&hugetlb_lock);
1452         }
1453
1454         return 0;
1455 }
1456
1457 #endif /* CONFIG_SYSCTL */
1458
1459 void hugetlb_report_meminfo(struct seq_file *m)
1460 {
1461         struct hstate *h = &default_hstate;
1462         seq_printf(m,
1463                         "HugePages_Total:   %5lu\n"
1464                         "HugePages_Free:    %5lu\n"
1465                         "HugePages_Rsvd:    %5lu\n"
1466                         "HugePages_Surp:    %5lu\n"
1467                         "Hugepagesize:   %8lu kB\n",
1468                         h->nr_huge_pages,
1469                         h->free_huge_pages,
1470                         h->resv_huge_pages,
1471                         h->surplus_huge_pages,
1472                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1473 }
1474
1475 int hugetlb_report_node_meminfo(int nid, char *buf)
1476 {
1477         struct hstate *h = &default_hstate;
1478         return sprintf(buf,
1479                 "Node %d HugePages_Total: %5u\n"
1480                 "Node %d HugePages_Free:  %5u\n"
1481                 "Node %d HugePages_Surp:  %5u\n",
1482                 nid, h->nr_huge_pages_node[nid],
1483                 nid, h->free_huge_pages_node[nid],
1484                 nid, h->surplus_huge_pages_node[nid]);
1485 }
1486
1487 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1488 unsigned long hugetlb_total_pages(void)
1489 {
1490         struct hstate *h = &default_hstate;
1491         return h->nr_huge_pages * pages_per_huge_page(h);
1492 }
1493
1494 static int hugetlb_acct_memory(struct hstate *h, long delta)
1495 {
1496         int ret = -ENOMEM;
1497
1498         spin_lock(&hugetlb_lock);
1499         /*
1500          * When cpuset is configured, it breaks the strict hugetlb page
1501          * reservation as the accounting is done on a global variable. Such
1502          * reservation is completely rubbish in the presence of cpuset because
1503          * the reservation is not checked against page availability for the
1504          * current cpuset. Application can still potentially OOM'ed by kernel
1505          * with lack of free htlb page in cpuset that the task is in.
1506          * Attempt to enforce strict accounting with cpuset is almost
1507          * impossible (or too ugly) because cpuset is too fluid that
1508          * task or memory node can be dynamically moved between cpusets.
1509          *
1510          * The change of semantics for shared hugetlb mapping with cpuset is
1511          * undesirable. However, in order to preserve some of the semantics,
1512          * we fall back to check against current free page availability as
1513          * a best attempt and hopefully to minimize the impact of changing
1514          * semantics that cpuset has.
1515          */
1516         if (delta > 0) {
1517                 if (gather_surplus_pages(h, delta) < 0)
1518                         goto out;
1519
1520                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1521                         return_unused_surplus_pages(h, delta);
1522                         goto out;
1523                 }
1524         }
1525
1526         ret = 0;
1527         if (delta < 0)
1528                 return_unused_surplus_pages(h, (unsigned long) -delta);
1529
1530 out:
1531         spin_unlock(&hugetlb_lock);
1532         return ret;
1533 }
1534
1535 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1536 {
1537         struct resv_map *reservations = vma_resv_map(vma);
1538
1539         /*
1540          * This new VMA should share its siblings reservation map if present.
1541          * The VMA will only ever have a valid reservation map pointer where
1542          * it is being copied for another still existing VMA.  As that VMA
1543          * has a reference to the reservation map it cannot dissappear until
1544          * after this open call completes.  It is therefore safe to take a
1545          * new reference here without additional locking.
1546          */
1547         if (reservations)
1548                 kref_get(&reservations->refs);
1549 }
1550
1551 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1552 {
1553         struct hstate *h = hstate_vma(vma);
1554         struct resv_map *reservations = vma_resv_map(vma);
1555         unsigned long reserve;
1556         unsigned long start;
1557         unsigned long end;
1558
1559         if (reservations) {
1560                 start = vma_hugecache_offset(h, vma, vma->vm_start);
1561                 end = vma_hugecache_offset(h, vma, vma->vm_end);
1562
1563                 reserve = (end - start) -
1564                         region_count(&reservations->regions, start, end);
1565
1566                 kref_put(&reservations->refs, resv_map_release);
1567
1568                 if (reserve) {
1569                         hugetlb_acct_memory(h, -reserve);
1570                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1571                 }
1572         }
1573 }
1574
1575 /*
1576  * We cannot handle pagefaults against hugetlb pages at all.  They cause
1577  * handle_mm_fault() to try to instantiate regular-sized pages in the
1578  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
1579  * this far.
1580  */
1581 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1582 {
1583         BUG();
1584         return 0;
1585 }
1586
1587 struct vm_operations_struct hugetlb_vm_ops = {
1588         .fault = hugetlb_vm_op_fault,
1589         .open = hugetlb_vm_op_open,
1590         .close = hugetlb_vm_op_close,
1591 };
1592
1593 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1594                                 int writable)
1595 {
1596         pte_t entry;
1597
1598         if (writable) {
1599                 entry =
1600                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1601         } else {
1602                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1603         }
1604         entry = pte_mkyoung(entry);
1605         entry = pte_mkhuge(entry);
1606
1607         return entry;
1608 }
1609
1610 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1611                                    unsigned long address, pte_t *ptep)
1612 {
1613         pte_t entry;
1614
1615         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1616         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1617                 update_mmu_cache(vma, address, entry);
1618         }
1619 }
1620
1621
1622 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1623                             struct vm_area_struct *vma)
1624 {
1625         pte_t *src_pte, *dst_pte, entry;
1626         struct page *ptepage;
1627         unsigned long addr;
1628         int cow;
1629         struct hstate *h = hstate_vma(vma);
1630         unsigned long sz = huge_page_size(h);
1631
1632         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1633
1634         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1635                 src_pte = huge_pte_offset(src, addr);
1636                 if (!src_pte)
1637                         continue;
1638                 dst_pte = huge_pte_alloc(dst, addr, sz);
1639                 if (!dst_pte)
1640                         goto nomem;
1641
1642                 /* If the pagetables are shared don't copy or take references */
1643                 if (dst_pte == src_pte)
1644                         continue;
1645
1646                 spin_lock(&dst->page_table_lock);
1647                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1648                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1649                         if (cow)
1650                                 huge_ptep_set_wrprotect(src, addr, src_pte);
1651                         entry = huge_ptep_get(src_pte);
1652                         ptepage = pte_page(entry);
1653                         get_page(ptepage);
1654                         set_huge_pte_at(dst, addr, dst_pte, entry);
1655                 }
1656                 spin_unlock(&src->page_table_lock);
1657                 spin_unlock(&dst->page_table_lock);
1658         }
1659         return 0;
1660
1661 nomem:
1662         return -ENOMEM;
1663 }
1664
1665 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1666                             unsigned long end, struct page *ref_page)
1667 {
1668         struct mm_struct *mm = vma->vm_mm;
1669         unsigned long address;
1670         pte_t *ptep;
1671         pte_t pte;
1672         struct page *page;
1673         struct page *tmp;
1674         struct hstate *h = hstate_vma(vma);
1675         unsigned long sz = huge_page_size(h);
1676
1677         /*
1678          * A page gathering list, protected by per file i_mmap_lock. The
1679          * lock is used to avoid list corruption from multiple unmapping
1680          * of the same page since we are using page->lru.
1681          */
1682         LIST_HEAD(page_list);
1683
1684         WARN_ON(!is_vm_hugetlb_page(vma));
1685         BUG_ON(start & ~huge_page_mask(h));
1686         BUG_ON(end & ~huge_page_mask(h));
1687
1688         mmu_notifier_invalidate_range_start(mm, start, end);
1689         spin_lock(&mm->page_table_lock);
1690         for (address = start; address < end; address += sz) {
1691                 ptep = huge_pte_offset(mm, address);
1692                 if (!ptep)
1693                         continue;
1694
1695                 if (huge_pmd_unshare(mm, &address, ptep))
1696                         continue;
1697
1698                 /*
1699                  * If a reference page is supplied, it is because a specific
1700                  * page is being unmapped, not a range. Ensure the page we
1701                  * are about to unmap is the actual page of interest.
1702                  */
1703                 if (ref_page) {
1704                         pte = huge_ptep_get(ptep);
1705                         if (huge_pte_none(pte))
1706                                 continue;
1707                         page = pte_page(pte);
1708                         if (page != ref_page)
1709                                 continue;
1710
1711                         /*
1712                          * Mark the VMA as having unmapped its page so that
1713                          * future faults in this VMA will fail rather than
1714                          * looking like data was lost
1715                          */
1716                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1717                 }
1718
1719                 pte = huge_ptep_get_and_clear(mm, address, ptep);
1720                 if (huge_pte_none(pte))
1721                         continue;
1722
1723                 page = pte_page(pte);
1724                 if (pte_dirty(pte))
1725                         set_page_dirty(page);
1726                 list_add(&page->lru, &page_list);
1727         }
1728         spin_unlock(&mm->page_table_lock);
1729         flush_tlb_range(vma, start, end);
1730         mmu_notifier_invalidate_range_end(mm, start, end);
1731         list_for_each_entry_safe(page, tmp, &page_list, lru) {
1732                 list_del(&page->lru);
1733                 put_page(page);
1734         }
1735 }
1736
1737 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1738                           unsigned long end, struct page *ref_page)
1739 {
1740         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1741         __unmap_hugepage_range(vma, start, end, ref_page);
1742         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1743 }
1744
1745 /*
1746  * This is called when the original mapper is failing to COW a MAP_PRIVATE
1747  * mappping it owns the reserve page for. The intention is to unmap the page
1748  * from other VMAs and let the children be SIGKILLed if they are faulting the
1749  * same region.
1750  */
1751 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1752                                 struct page *page, unsigned long address)
1753 {
1754         struct vm_area_struct *iter_vma;
1755         struct address_space *mapping;
1756         struct prio_tree_iter iter;
1757         pgoff_t pgoff;
1758
1759         /*
1760          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1761          * from page cache lookup which is in HPAGE_SIZE units.
1762          */
1763         address = address & huge_page_mask(hstate_vma(vma));
1764         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1765                 + (vma->vm_pgoff >> PAGE_SHIFT);
1766         mapping = (struct address_space *)page_private(page);
1767
1768         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1769                 /* Do not unmap the current VMA */
1770                 if (iter_vma == vma)
1771                         continue;
1772
1773                 /*
1774                  * Unmap the page from other VMAs without their own reserves.
1775                  * They get marked to be SIGKILLed if they fault in these
1776                  * areas. This is because a future no-page fault on this VMA
1777                  * could insert a zeroed page instead of the data existing
1778                  * from the time of fork. This would look like data corruption
1779                  */
1780                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1781                         unmap_hugepage_range(iter_vma,
1782                                 address, address + HPAGE_SIZE,
1783                                 page);
1784         }
1785
1786         return 1;
1787 }
1788
1789 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1790                         unsigned long address, pte_t *ptep, pte_t pte,
1791                         struct page *pagecache_page)
1792 {
1793         struct hstate *h = hstate_vma(vma);
1794         struct page *old_page, *new_page;
1795         int avoidcopy;
1796         int outside_reserve = 0;
1797
1798         old_page = pte_page(pte);
1799
1800 retry_avoidcopy:
1801         /* If no-one else is actually using this page, avoid the copy
1802          * and just make the page writable */
1803         avoidcopy = (page_count(old_page) == 1);
1804         if (avoidcopy) {
1805                 set_huge_ptep_writable(vma, address, ptep);
1806                 return 0;
1807         }
1808
1809         /*
1810          * If the process that created a MAP_PRIVATE mapping is about to
1811          * perform a COW due to a shared page count, attempt to satisfy
1812          * the allocation without using the existing reserves. The pagecache
1813          * page is used to determine if the reserve at this address was
1814          * consumed or not. If reserves were used, a partial faulted mapping
1815          * at the time of fork() could consume its reserves on COW instead
1816          * of the full address range.
1817          */
1818         if (!(vma->vm_flags & VM_SHARED) &&
1819                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1820                         old_page != pagecache_page)
1821                 outside_reserve = 1;
1822
1823         page_cache_get(old_page);
1824         new_page = alloc_huge_page(vma, address, outside_reserve);
1825
1826         if (IS_ERR(new_page)) {
1827                 page_cache_release(old_page);
1828
1829                 /*
1830                  * If a process owning a MAP_PRIVATE mapping fails to COW,
1831                  * it is due to references held by a child and an insufficient
1832                  * huge page pool. To guarantee the original mappers
1833                  * reliability, unmap the page from child processes. The child
1834                  * may get SIGKILLed if it later faults.
1835                  */
1836                 if (outside_reserve) {
1837                         BUG_ON(huge_pte_none(pte));
1838                         if (unmap_ref_private(mm, vma, old_page, address)) {
1839                                 BUG_ON(page_count(old_page) != 1);
1840                                 BUG_ON(huge_pte_none(pte));
1841                                 goto retry_avoidcopy;
1842                         }
1843                         WARN_ON_ONCE(1);
1844                 }
1845
1846                 return -PTR_ERR(new_page);
1847         }
1848
1849         spin_unlock(&mm->page_table_lock);
1850         copy_huge_page(new_page, old_page, address, vma);
1851         __SetPageUptodate(new_page);
1852         spin_lock(&mm->page_table_lock);
1853
1854         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1855         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1856                 /* Break COW */
1857                 huge_ptep_clear_flush(vma, address, ptep);
1858                 set_huge_pte_at(mm, address, ptep,
1859                                 make_huge_pte(vma, new_page, 1));
1860                 /* Make the old page be freed below */
1861                 new_page = old_page;
1862         }
1863         page_cache_release(new_page);
1864         page_cache_release(old_page);
1865         return 0;
1866 }
1867
1868 /* Return the pagecache page at a given address within a VMA */
1869 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1870                         struct vm_area_struct *vma, unsigned long address)
1871 {
1872         struct address_space *mapping;
1873         pgoff_t idx;
1874
1875         mapping = vma->vm_file->f_mapping;
1876         idx = vma_hugecache_offset(h, vma, address);
1877
1878         return find_lock_page(mapping, idx);
1879 }
1880
1881 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1882                         unsigned long address, pte_t *ptep, int write_access)
1883 {
1884         struct hstate *h = hstate_vma(vma);
1885         int ret = VM_FAULT_SIGBUS;
1886         pgoff_t idx;
1887         unsigned long size;
1888         struct page *page;
1889         struct address_space *mapping;
1890         pte_t new_pte;
1891
1892         /*
1893          * Currently, we are forced to kill the process in the event the
1894          * original mapper has unmapped pages from the child due to a failed
1895          * COW. Warn that such a situation has occured as it may not be obvious
1896          */
1897         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1898                 printk(KERN_WARNING
1899                         "PID %d killed due to inadequate hugepage pool\n",
1900                         current->pid);
1901                 return ret;
1902         }
1903
1904         mapping = vma->vm_file->f_mapping;
1905         idx = vma_hugecache_offset(h, vma, address);
1906
1907         /*
1908          * Use page lock to guard against racing truncation
1909          * before we get page_table_lock.
1910          */
1911 retry:
1912         page = find_lock_page(mapping, idx);
1913         if (!page) {
1914                 size = i_size_read(mapping->host) >> huge_page_shift(h);
1915                 if (idx >= size)
1916                         goto out;
1917                 page = alloc_huge_page(vma, address, 0);
1918                 if (IS_ERR(page)) {
1919                         ret = -PTR_ERR(page);
1920                         goto out;
1921                 }
1922                 clear_huge_page(page, address, huge_page_size(h));
1923                 __SetPageUptodate(page);
1924
1925                 if (vma->vm_flags & VM_SHARED) {
1926                         int err;
1927                         struct inode *inode = mapping->host;
1928
1929                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1930                         if (err) {
1931                                 put_page(page);
1932                                 if (err == -EEXIST)
1933                                         goto retry;
1934                                 goto out;
1935                         }
1936
1937                         spin_lock(&inode->i_lock);
1938                         inode->i_blocks += blocks_per_huge_page(h);
1939                         spin_unlock(&inode->i_lock);
1940                 } else
1941                         lock_page(page);
1942         }
1943
1944         /*
1945          * If we are going to COW a private mapping later, we examine the
1946          * pending reservations for this page now. This will ensure that
1947          * any allocations necessary to record that reservation occur outside
1948          * the spinlock.
1949          */
1950         if (write_access && !(vma->vm_flags & VM_SHARED))
1951                 if (vma_needs_reservation(h, vma, address) < 0) {
1952                         ret = VM_FAULT_OOM;
1953                         goto backout_unlocked;
1954                 }
1955
1956         spin_lock(&mm->page_table_lock);
1957         size = i_size_read(mapping->host) >> huge_page_shift(h);
1958         if (idx >= size)
1959                 goto backout;
1960
1961         ret = 0;
1962         if (!huge_pte_none(huge_ptep_get(ptep)))
1963                 goto backout;
1964
1965         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
1966                                 && (vma->vm_flags & VM_SHARED)));
1967         set_huge_pte_at(mm, address, ptep, new_pte);
1968
1969         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1970                 /* Optimization, do the COW without a second fault */
1971                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
1972         }
1973
1974         spin_unlock(&mm->page_table_lock);
1975         unlock_page(page);
1976 out:
1977         return ret;
1978
1979 backout:
1980         spin_unlock(&mm->page_table_lock);
1981 backout_unlocked:
1982         unlock_page(page);
1983         put_page(page);
1984         goto out;
1985 }
1986
1987 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1988                         unsigned long address, int write_access)
1989 {
1990         pte_t *ptep;
1991         pte_t entry;
1992         int ret;
1993         struct page *pagecache_page = NULL;
1994         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
1995         struct hstate *h = hstate_vma(vma);
1996
1997         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
1998         if (!ptep)
1999                 return VM_FAULT_OOM;
2000
2001         /*
2002          * Serialize hugepage allocation and instantiation, so that we don't
2003          * get spurious allocation failures if two CPUs race to instantiate
2004          * the same page in the page cache.
2005          */
2006         mutex_lock(&hugetlb_instantiation_mutex);
2007         entry = huge_ptep_get(ptep);
2008         if (huge_pte_none(entry)) {
2009                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
2010                 goto out_mutex;
2011         }
2012
2013         ret = 0;
2014
2015         /*
2016          * If we are going to COW the mapping later, we examine the pending
2017          * reservations for this page now. This will ensure that any
2018          * allocations necessary to record that reservation occur outside the
2019          * spinlock. For private mappings, we also lookup the pagecache
2020          * page now as it is used to determine if a reservation has been
2021          * consumed.
2022          */
2023         if (write_access && !pte_write(entry)) {
2024                 if (vma_needs_reservation(h, vma, address) < 0) {
2025                         ret = VM_FAULT_OOM;
2026                         goto out_mutex;
2027                 }
2028
2029                 if (!(vma->vm_flags & VM_SHARED))
2030                         pagecache_page = hugetlbfs_pagecache_page(h,
2031                                                                 vma, address);
2032         }
2033
2034         spin_lock(&mm->page_table_lock);
2035         /* Check for a racing update before calling hugetlb_cow */
2036         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2037                 goto out_page_table_lock;
2038
2039
2040         if (write_access) {
2041                 if (!pte_write(entry)) {
2042                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2043                                                         pagecache_page);
2044                         goto out_page_table_lock;
2045                 }
2046                 entry = pte_mkdirty(entry);
2047         }
2048         entry = pte_mkyoung(entry);
2049         if (huge_ptep_set_access_flags(vma, address, ptep, entry, write_access))
2050                 update_mmu_cache(vma, address, entry);
2051
2052 out_page_table_lock:
2053         spin_unlock(&mm->page_table_lock);
2054
2055         if (pagecache_page) {
2056                 unlock_page(pagecache_page);
2057                 put_page(pagecache_page);
2058         }
2059
2060 out_mutex:
2061         mutex_unlock(&hugetlb_instantiation_mutex);
2062
2063         return ret;
2064 }
2065
2066 /* Can be overriden by architectures */
2067 __attribute__((weak)) struct page *
2068 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2069                pud_t *pud, int write)
2070 {
2071         BUG();
2072         return NULL;
2073 }
2074
2075 static int huge_zeropage_ok(pte_t *ptep, int write, int shared)
2076 {
2077         if (!ptep || write || shared)
2078                 return 0;
2079         else
2080                 return huge_pte_none(huge_ptep_get(ptep));
2081 }
2082
2083 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2084                         struct page **pages, struct vm_area_struct **vmas,
2085                         unsigned long *position, int *length, int i,
2086                         int write)
2087 {
2088         unsigned long pfn_offset;
2089         unsigned long vaddr = *position;
2090         int remainder = *length;
2091         struct hstate *h = hstate_vma(vma);
2092         int zeropage_ok = 0;
2093         int shared = vma->vm_flags & VM_SHARED;
2094
2095         spin_lock(&mm->page_table_lock);
2096         while (vaddr < vma->vm_end && remainder) {
2097                 pte_t *pte;
2098                 struct page *page;
2099
2100                 /*
2101                  * Some archs (sparc64, sh*) have multiple pte_ts to
2102                  * each hugepage.  We have to make * sure we get the
2103                  * first, for the page indexing below to work.
2104                  */
2105                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2106                 if (huge_zeropage_ok(pte, write, shared))
2107                         zeropage_ok = 1;
2108
2109                 if (!pte ||
2110                     (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) ||
2111                     (write && !pte_write(huge_ptep_get(pte)))) {
2112                         int ret;
2113
2114                         spin_unlock(&mm->page_table_lock);
2115                         ret = hugetlb_fault(mm, vma, vaddr, write);
2116                         spin_lock(&mm->page_table_lock);
2117                         if (!(ret & VM_FAULT_ERROR))
2118                                 continue;
2119
2120                         remainder = 0;
2121                         if (!i)
2122                                 i = -EFAULT;
2123                         break;
2124                 }
2125
2126                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2127                 page = pte_page(huge_ptep_get(pte));
2128 same_page:
2129                 if (pages) {
2130                         if (zeropage_ok)
2131                                 pages[i] = ZERO_PAGE(0);
2132                         else
2133                                 pages[i] = page + pfn_offset;
2134                         get_page(pages[i]);
2135                 }
2136
2137                 if (vmas)
2138                         vmas[i] = vma;
2139
2140                 vaddr += PAGE_SIZE;
2141                 ++pfn_offset;
2142                 --remainder;
2143                 ++i;
2144                 if (vaddr < vma->vm_end && remainder &&
2145                                 pfn_offset < pages_per_huge_page(h)) {
2146                         /*
2147                          * We use pfn_offset to avoid touching the pageframes
2148                          * of this compound page.
2149                          */
2150                         goto same_page;
2151                 }
2152         }
2153         spin_unlock(&mm->page_table_lock);
2154         *length = remainder;
2155         *position = vaddr;
2156
2157         return i;
2158 }
2159
2160 void hugetlb_change_protection(struct vm_area_struct *vma,
2161                 unsigned long address, unsigned long end, pgprot_t newprot)
2162 {
2163         struct mm_struct *mm = vma->vm_mm;
2164         unsigned long start = address;
2165         pte_t *ptep;
2166         pte_t pte;
2167         struct hstate *h = hstate_vma(vma);
2168
2169         BUG_ON(address >= end);
2170         flush_cache_range(vma, address, end);
2171
2172         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2173         spin_lock(&mm->page_table_lock);
2174         for (; address < end; address += huge_page_size(h)) {
2175                 ptep = huge_pte_offset(mm, address);
2176                 if (!ptep)
2177                         continue;
2178                 if (huge_pmd_unshare(mm, &address, ptep))
2179                         continue;
2180                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2181                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2182                         pte = pte_mkhuge(pte_modify(pte, newprot));
2183                         set_huge_pte_at(mm, address, ptep, pte);
2184                 }
2185         }
2186         spin_unlock(&mm->page_table_lock);
2187         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2188
2189         flush_tlb_range(vma, start, end);
2190 }
2191
2192 int hugetlb_reserve_pages(struct inode *inode,
2193                                         long from, long to,
2194                                         struct vm_area_struct *vma)
2195 {
2196         long ret, chg;
2197         struct hstate *h = hstate_inode(inode);
2198
2199         if (vma && vma->vm_flags & VM_NORESERVE)
2200                 return 0;
2201
2202         /*
2203          * Shared mappings base their reservation on the number of pages that
2204          * are already allocated on behalf of the file. Private mappings need
2205          * to reserve the full area even if read-only as mprotect() may be
2206          * called to make the mapping read-write. Assume !vma is a shm mapping
2207          */
2208         if (!vma || vma->vm_flags & VM_SHARED)
2209                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2210         else {
2211                 struct resv_map *resv_map = resv_map_alloc();
2212                 if (!resv_map)
2213                         return -ENOMEM;
2214
2215                 chg = to - from;
2216
2217                 set_vma_resv_map(vma, resv_map);
2218                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2219         }
2220
2221         if (chg < 0)
2222                 return chg;
2223
2224         if (hugetlb_get_quota(inode->i_mapping, chg))
2225                 return -ENOSPC;
2226         ret = hugetlb_acct_memory(h, chg);
2227         if (ret < 0) {
2228                 hugetlb_put_quota(inode->i_mapping, chg);
2229                 return ret;
2230         }
2231         if (!vma || vma->vm_flags & VM_SHARED)
2232                 region_add(&inode->i_mapping->private_list, from, to);
2233         return 0;
2234 }
2235
2236 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2237 {
2238         struct hstate *h = hstate_inode(inode);
2239         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2240
2241         spin_lock(&inode->i_lock);
2242         inode->i_blocks -= blocks_per_huge_page(h);
2243         spin_unlock(&inode->i_lock);
2244
2245         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2246         hugetlb_acct_memory(h, -(chg - freed));
2247 }