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