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