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