mm: report the MMU pagesize in /proc/pid/smaps
[linux-2.6.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) William Irwin, April 2004
4  */
5 #include <linux/gfp.h>
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.h>
9 #include <linux/mm.h>
10 #include <linux/seq_file.h>
11 #include <linux/sysctl.h>
12 #include <linux/highmem.h>
13 #include <linux/mmu_notifier.h>
14 #include <linux/nodemask.h>
15 #include <linux/pagemap.h>
16 #include <linux/mempolicy.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21
22 #include <asm/page.h>
23 #include <asm/pgtable.h>
24 #include <asm/io.h>
25
26 #include <linux/hugetlb.h>
27 #include "internal.h"
28
29 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
30 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
31 unsigned long hugepages_treat_as_movable;
32
33 static int max_hstate;
34 unsigned int default_hstate_idx;
35 struct hstate hstates[HUGE_MAX_HSTATE];
36
37 __initdata LIST_HEAD(huge_boot_pages);
38
39 /* for command line parsing */
40 static struct hstate * __initdata parsed_hstate;
41 static unsigned long __initdata default_hstate_max_huge_pages;
42 static unsigned long __initdata default_hstate_size;
43
44 #define for_each_hstate(h) \
45         for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
46
47 /*
48  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
49  */
50 static DEFINE_SPINLOCK(hugetlb_lock);
51
52 /*
53  * Region tracking -- allows tracking of reservations and instantiated pages
54  *                    across the pages in a mapping.
55  *
56  * The region data structures are protected by a combination of the mmap_sem
57  * and the hugetlb_instantion_mutex.  To access or modify a region the caller
58  * must either hold the mmap_sem for write, or the mmap_sem for read and
59  * the hugetlb_instantiation mutex:
60  *
61  *      down_write(&mm->mmap_sem);
62  * or
63  *      down_read(&mm->mmap_sem);
64  *      mutex_lock(&hugetlb_instantiation_mutex);
65  */
66 struct file_region {
67         struct list_head link;
68         long from;
69         long to;
70 };
71
72 static long region_add(struct list_head *head, long f, long t)
73 {
74         struct file_region *rg, *nrg, *trg;
75
76         /* Locate the region we are either in or before. */
77         list_for_each_entry(rg, head, link)
78                 if (f <= rg->to)
79                         break;
80
81         /* Round our left edge to the current segment if it encloses us. */
82         if (f > rg->from)
83                 f = rg->from;
84
85         /* Check for and consume any regions we now overlap with. */
86         nrg = rg;
87         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
88                 if (&rg->link == head)
89                         break;
90                 if (rg->from > t)
91                         break;
92
93                 /* If this area reaches higher then extend our area to
94                  * include it completely.  If this is not the first area
95                  * which we intend to reuse, free it. */
96                 if (rg->to > t)
97                         t = rg->to;
98                 if (rg != nrg) {
99                         list_del(&rg->link);
100                         kfree(rg);
101                 }
102         }
103         nrg->from = f;
104         nrg->to = t;
105         return 0;
106 }
107
108 static long region_chg(struct list_head *head, long f, long t)
109 {
110         struct file_region *rg, *nrg;
111         long chg = 0;
112
113         /* Locate the region we are before or in. */
114         list_for_each_entry(rg, head, link)
115                 if (f <= rg->to)
116                         break;
117
118         /* If we are below the current region then a new region is required.
119          * Subtle, allocate a new region at the position but make it zero
120          * size such that we can guarantee to record the reservation. */
121         if (&rg->link == head || t < rg->from) {
122                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
123                 if (!nrg)
124                         return -ENOMEM;
125                 nrg->from = f;
126                 nrg->to   = f;
127                 INIT_LIST_HEAD(&nrg->link);
128                 list_add(&nrg->link, rg->link.prev);
129
130                 return t - f;
131         }
132
133         /* Round our left edge to the current segment if it encloses us. */
134         if (f > rg->from)
135                 f = rg->from;
136         chg = t - f;
137
138         /* Check for and consume any regions we now overlap with. */
139         list_for_each_entry(rg, rg->link.prev, link) {
140                 if (&rg->link == head)
141                         break;
142                 if (rg->from > t)
143                         return chg;
144
145                 /* We overlap with this area, if it extends futher than
146                  * us then we must extend ourselves.  Account for its
147                  * existing reservation. */
148                 if (rg->to > t) {
149                         chg += rg->to - t;
150                         t = rg->to;
151                 }
152                 chg -= rg->to - rg->from;
153         }
154         return chg;
155 }
156
157 static long region_truncate(struct list_head *head, long end)
158 {
159         struct file_region *rg, *trg;
160         long chg = 0;
161
162         /* Locate the region we are either in or before. */
163         list_for_each_entry(rg, head, link)
164                 if (end <= rg->to)
165                         break;
166         if (&rg->link == head)
167                 return 0;
168
169         /* If we are in the middle of a region then adjust it. */
170         if (end > rg->from) {
171                 chg = rg->to - end;
172                 rg->to = end;
173                 rg = list_entry(rg->link.next, typeof(*rg), link);
174         }
175
176         /* Drop any remaining regions. */
177         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
178                 if (&rg->link == head)
179                         break;
180                 chg += rg->to - rg->from;
181                 list_del(&rg->link);
182                 kfree(rg);
183         }
184         return chg;
185 }
186
187 static long region_count(struct list_head *head, long f, long t)
188 {
189         struct file_region *rg;
190         long chg = 0;
191
192         /* Locate each segment we overlap with, and count that overlap. */
193         list_for_each_entry(rg, head, link) {
194                 int seg_from;
195                 int seg_to;
196
197                 if (rg->to <= f)
198                         continue;
199                 if (rg->from >= t)
200                         break;
201
202                 seg_from = max(rg->from, f);
203                 seg_to = min(rg->to, t);
204
205                 chg += seg_to - seg_from;
206         }
207
208         return chg;
209 }
210
211 /*
212  * Convert the address within this vma to the page offset within
213  * the mapping, in pagecache page units; huge pages here.
214  */
215 static pgoff_t vma_hugecache_offset(struct hstate *h,
216                         struct vm_area_struct *vma, unsigned long address)
217 {
218         return ((address - vma->vm_start) >> huge_page_shift(h)) +
219                         (vma->vm_pgoff >> huge_page_order(h));
220 }
221
222 /*
223  * Return the size of the pages allocated when backing a VMA. In the majority
224  * cases this will be same size as used by the page table entries.
225  */
226 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
227 {
228         struct hstate *hstate;
229
230         if (!is_vm_hugetlb_page(vma))
231                 return PAGE_SIZE;
232
233         hstate = hstate_vma(vma);
234
235         return 1UL << (hstate->order + PAGE_SHIFT);
236 }
237
238 /*
239  * Return the page size being used by the MMU to back a VMA. In the majority
240  * of cases, the page size used by the kernel matches the MMU size. On
241  * architectures where it differs, an architecture-specific version of this
242  * function is required.
243  */
244 #ifndef vma_mmu_pagesize
245 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
246 {
247         return vma_kernel_pagesize(vma);
248 }
249 #endif
250
251 /*
252  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
253  * bits of the reservation map pointer, which are always clear due to
254  * alignment.
255  */
256 #define HPAGE_RESV_OWNER    (1UL << 0)
257 #define HPAGE_RESV_UNMAPPED (1UL << 1)
258 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
259
260 /*
261  * These helpers are used to track how many pages are reserved for
262  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
263  * is guaranteed to have their future faults succeed.
264  *
265  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
266  * the reserve counters are updated with the hugetlb_lock held. It is safe
267  * to reset the VMA at fork() time as it is not in use yet and there is no
268  * chance of the global counters getting corrupted as a result of the values.
269  *
270  * The private mapping reservation is represented in a subtly different
271  * manner to a shared mapping.  A shared mapping has a region map associated
272  * with the underlying file, this region map represents the backing file
273  * pages which have ever had a reservation assigned which this persists even
274  * after the page is instantiated.  A private mapping has a region map
275  * associated with the original mmap which is attached to all VMAs which
276  * reference it, this region map represents those offsets which have consumed
277  * reservation ie. where pages have been instantiated.
278  */
279 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
280 {
281         return (unsigned long)vma->vm_private_data;
282 }
283
284 static void set_vma_private_data(struct vm_area_struct *vma,
285                                                         unsigned long value)
286 {
287         vma->vm_private_data = (void *)value;
288 }
289
290 struct resv_map {
291         struct kref refs;
292         struct list_head regions;
293 };
294
295 static struct resv_map *resv_map_alloc(void)
296 {
297         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
298         if (!resv_map)
299                 return NULL;
300
301         kref_init(&resv_map->refs);
302         INIT_LIST_HEAD(&resv_map->regions);
303
304         return resv_map;
305 }
306
307 static void resv_map_release(struct kref *ref)
308 {
309         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
310
311         /* Clear out any active regions before we release the map. */
312         region_truncate(&resv_map->regions, 0);
313         kfree(resv_map);
314 }
315
316 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
317 {
318         VM_BUG_ON(!is_vm_hugetlb_page(vma));
319         if (!(vma->vm_flags & VM_SHARED))
320                 return (struct resv_map *)(get_vma_private_data(vma) &
321                                                         ~HPAGE_RESV_MASK);
322         return NULL;
323 }
324
325 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
326 {
327         VM_BUG_ON(!is_vm_hugetlb_page(vma));
328         VM_BUG_ON(vma->vm_flags & VM_SHARED);
329
330         set_vma_private_data(vma, (get_vma_private_data(vma) &
331                                 HPAGE_RESV_MASK) | (unsigned long)map);
332 }
333
334 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
335 {
336         VM_BUG_ON(!is_vm_hugetlb_page(vma));
337         VM_BUG_ON(vma->vm_flags & VM_SHARED);
338
339         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
340 }
341
342 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
343 {
344         VM_BUG_ON(!is_vm_hugetlb_page(vma));
345
346         return (get_vma_private_data(vma) & flag) != 0;
347 }
348
349 /* Decrement the reserved pages in the hugepage pool by one */
350 static void decrement_hugepage_resv_vma(struct hstate *h,
351                         struct vm_area_struct *vma)
352 {
353         if (vma->vm_flags & VM_NORESERVE)
354                 return;
355
356         if (vma->vm_flags & VM_SHARED) {
357                 /* Shared mappings always use reserves */
358                 h->resv_huge_pages--;
359         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
360                 /*
361                  * Only the process that called mmap() has reserves for
362                  * private mappings.
363                  */
364                 h->resv_huge_pages--;
365         }
366 }
367
368 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
369 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
370 {
371         VM_BUG_ON(!is_vm_hugetlb_page(vma));
372         if (!(vma->vm_flags & VM_SHARED))
373                 vma->vm_private_data = (void *)0;
374 }
375
376 /* Returns true if the VMA has associated reserve pages */
377 static int vma_has_reserves(struct vm_area_struct *vma)
378 {
379         if (vma->vm_flags & VM_SHARED)
380                 return 1;
381         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
382                 return 1;
383         return 0;
384 }
385
386 static void clear_gigantic_page(struct page *page,
387                         unsigned long addr, unsigned long sz)
388 {
389         int i;
390         struct page *p = page;
391
392         might_sleep();
393         for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
394                 cond_resched();
395                 clear_user_highpage(p, addr + i * PAGE_SIZE);
396         }
397 }
398 static void clear_huge_page(struct page *page,
399                         unsigned long addr, unsigned long sz)
400 {
401         int i;
402
403         if (unlikely(sz > MAX_ORDER_NR_PAGES))
404                 return clear_gigantic_page(page, addr, sz);
405
406         might_sleep();
407         for (i = 0; i < sz/PAGE_SIZE; i++) {
408                 cond_resched();
409                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
410         }
411 }
412
413 static void copy_gigantic_page(struct page *dst, struct page *src,
414                            unsigned long addr, struct vm_area_struct *vma)
415 {
416         int i;
417         struct hstate *h = hstate_vma(vma);
418         struct page *dst_base = dst;
419         struct page *src_base = src;
420         might_sleep();
421         for (i = 0; i < pages_per_huge_page(h); ) {
422                 cond_resched();
423                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
424
425                 i++;
426                 dst = mem_map_next(dst, dst_base, i);
427                 src = mem_map_next(src, src_base, i);
428         }
429 }
430 static void copy_huge_page(struct page *dst, struct page *src,
431                            unsigned long addr, struct vm_area_struct *vma)
432 {
433         int i;
434         struct hstate *h = hstate_vma(vma);
435
436         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES))
437                 return copy_gigantic_page(dst, src, addr, vma);
438
439         might_sleep();
440         for (i = 0; i < pages_per_huge_page(h); i++) {
441                 cond_resched();
442                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
443         }
444 }
445
446 static void enqueue_huge_page(struct hstate *h, struct page *page)
447 {
448         int nid = page_to_nid(page);
449         list_add(&page->lru, &h->hugepage_freelists[nid]);
450         h->free_huge_pages++;
451         h->free_huge_pages_node[nid]++;
452 }
453
454 static struct page *dequeue_huge_page(struct hstate *h)
455 {
456         int nid;
457         struct page *page = NULL;
458
459         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
460                 if (!list_empty(&h->hugepage_freelists[nid])) {
461                         page = list_entry(h->hugepage_freelists[nid].next,
462                                           struct page, lru);
463                         list_del(&page->lru);
464                         h->free_huge_pages--;
465                         h->free_huge_pages_node[nid]--;
466                         break;
467                 }
468         }
469         return page;
470 }
471
472 static struct page *dequeue_huge_page_vma(struct hstate *h,
473                                 struct vm_area_struct *vma,
474                                 unsigned long address, int avoid_reserve)
475 {
476         int nid;
477         struct page *page = NULL;
478         struct mempolicy *mpol;
479         nodemask_t *nodemask;
480         struct zonelist *zonelist = huge_zonelist(vma, address,
481                                         htlb_alloc_mask, &mpol, &nodemask);
482         struct zone *zone;
483         struct zoneref *z;
484
485         /*
486          * A child process with MAP_PRIVATE mappings created by their parent
487          * have no page reserves. This check ensures that reservations are
488          * not "stolen". The child may still get SIGKILLed
489          */
490         if (!vma_has_reserves(vma) &&
491                         h->free_huge_pages - h->resv_huge_pages == 0)
492                 return NULL;
493
494         /* If reserves cannot be used, ensure enough pages are in the pool */
495         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
496                 return NULL;
497
498         for_each_zone_zonelist_nodemask(zone, z, zonelist,
499                                                 MAX_NR_ZONES - 1, nodemask) {
500                 nid = zone_to_nid(zone);
501                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
502                     !list_empty(&h->hugepage_freelists[nid])) {
503                         page = list_entry(h->hugepage_freelists[nid].next,
504                                           struct page, lru);
505                         list_del(&page->lru);
506                         h->free_huge_pages--;
507                         h->free_huge_pages_node[nid]--;
508
509                         if (!avoid_reserve)
510                                 decrement_hugepage_resv_vma(h, vma);
511
512                         break;
513                 }
514         }
515         mpol_cond_put(mpol);
516         return page;
517 }
518
519 static void update_and_free_page(struct hstate *h, struct page *page)
520 {
521         int i;
522
523         VM_BUG_ON(h->order >= MAX_ORDER);
524
525         h->nr_huge_pages--;
526         h->nr_huge_pages_node[page_to_nid(page)]--;
527         for (i = 0; i < pages_per_huge_page(h); i++) {
528                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
529                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
530                                 1 << PG_private | 1<< PG_writeback);
531         }
532         set_compound_page_dtor(page, NULL);
533         set_page_refcounted(page);
534         arch_release_hugepage(page);
535         __free_pages(page, huge_page_order(h));
536 }
537
538 struct hstate *size_to_hstate(unsigned long size)
539 {
540         struct hstate *h;
541
542         for_each_hstate(h) {
543                 if (huge_page_size(h) == size)
544                         return h;
545         }
546         return NULL;
547 }
548
549 static void free_huge_page(struct page *page)
550 {
551         /*
552          * Can't pass hstate in here because it is called from the
553          * compound page destructor.
554          */
555         struct hstate *h = page_hstate(page);
556         int nid = page_to_nid(page);
557         struct address_space *mapping;
558
559         mapping = (struct address_space *) page_private(page);
560         set_page_private(page, 0);
561         BUG_ON(page_count(page));
562         INIT_LIST_HEAD(&page->lru);
563
564         spin_lock(&hugetlb_lock);
565         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
566                 update_and_free_page(h, page);
567                 h->surplus_huge_pages--;
568                 h->surplus_huge_pages_node[nid]--;
569         } else {
570                 enqueue_huge_page(h, page);
571         }
572         spin_unlock(&hugetlb_lock);
573         if (mapping)
574                 hugetlb_put_quota(mapping, 1);
575 }
576
577 /*
578  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
579  * balanced by operating on them in a round-robin fashion.
580  * Returns 1 if an adjustment was made.
581  */
582 static int adjust_pool_surplus(struct hstate *h, int delta)
583 {
584         static int prev_nid;
585         int nid = prev_nid;
586         int ret = 0;
587
588         VM_BUG_ON(delta != -1 && delta != 1);
589         do {
590                 nid = next_node(nid, node_online_map);
591                 if (nid == MAX_NUMNODES)
592                         nid = first_node(node_online_map);
593
594                 /* To shrink on this node, there must be a surplus page */
595                 if (delta < 0 && !h->surplus_huge_pages_node[nid])
596                         continue;
597                 /* Surplus cannot exceed the total number of pages */
598                 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
599                                                 h->nr_huge_pages_node[nid])
600                         continue;
601
602                 h->surplus_huge_pages += delta;
603                 h->surplus_huge_pages_node[nid] += delta;
604                 ret = 1;
605                 break;
606         } while (nid != prev_nid);
607
608         prev_nid = nid;
609         return ret;
610 }
611
612 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
613 {
614         set_compound_page_dtor(page, free_huge_page);
615         spin_lock(&hugetlb_lock);
616         h->nr_huge_pages++;
617         h->nr_huge_pages_node[nid]++;
618         spin_unlock(&hugetlb_lock);
619         put_page(page); /* free it into the hugepage allocator */
620 }
621
622 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
623 {
624         struct page *page;
625
626         if (h->order >= MAX_ORDER)
627                 return NULL;
628
629         page = alloc_pages_node(nid,
630                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
631                                                 __GFP_REPEAT|__GFP_NOWARN,
632                 huge_page_order(h));
633         if (page) {
634                 if (arch_prepare_hugepage(page)) {
635                         __free_pages(page, huge_page_order(h));
636                         return NULL;
637                 }
638                 prep_new_huge_page(h, page, nid);
639         }
640
641         return page;
642 }
643
644 /*
645  * Use a helper variable to find the next node and then
646  * copy it back to hugetlb_next_nid afterwards:
647  * otherwise there's a window in which a racer might
648  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
649  * But we don't need to use a spin_lock here: it really
650  * doesn't matter if occasionally a racer chooses the
651  * same nid as we do.  Move nid forward in the mask even
652  * if we just successfully allocated a hugepage so that
653  * the next caller gets hugepages on the next node.
654  */
655 static int hstate_next_node(struct hstate *h)
656 {
657         int next_nid;
658         next_nid = next_node(h->hugetlb_next_nid, node_online_map);
659         if (next_nid == MAX_NUMNODES)
660                 next_nid = first_node(node_online_map);
661         h->hugetlb_next_nid = next_nid;
662         return next_nid;
663 }
664
665 static int alloc_fresh_huge_page(struct hstate *h)
666 {
667         struct page *page;
668         int start_nid;
669         int next_nid;
670         int ret = 0;
671
672         start_nid = h->hugetlb_next_nid;
673
674         do {
675                 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
676                 if (page)
677                         ret = 1;
678                 next_nid = hstate_next_node(h);
679         } while (!page && h->hugetlb_next_nid != start_nid);
680
681         if (ret)
682                 count_vm_event(HTLB_BUDDY_PGALLOC);
683         else
684                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
685
686         return ret;
687 }
688
689 static struct page *alloc_buddy_huge_page(struct hstate *h,
690                         struct vm_area_struct *vma, unsigned long address)
691 {
692         struct page *page;
693         unsigned int nid;
694
695         if (h->order >= MAX_ORDER)
696                 return NULL;
697
698         /*
699          * Assume we will successfully allocate the surplus page to
700          * prevent racing processes from causing the surplus to exceed
701          * overcommit
702          *
703          * This however introduces a different race, where a process B
704          * tries to grow the static hugepage pool while alloc_pages() is
705          * called by process A. B will only examine the per-node
706          * counters in determining if surplus huge pages can be
707          * converted to normal huge pages in adjust_pool_surplus(). A
708          * won't be able to increment the per-node counter, until the
709          * lock is dropped by B, but B doesn't drop hugetlb_lock until
710          * no more huge pages can be converted from surplus to normal
711          * state (and doesn't try to convert again). Thus, we have a
712          * case where a surplus huge page exists, the pool is grown, and
713          * the surplus huge page still exists after, even though it
714          * should just have been converted to a normal huge page. This
715          * does not leak memory, though, as the hugepage will be freed
716          * once it is out of use. It also does not allow the counters to
717          * go out of whack in adjust_pool_surplus() as we don't modify
718          * the node values until we've gotten the hugepage and only the
719          * per-node value is checked there.
720          */
721         spin_lock(&hugetlb_lock);
722         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
723                 spin_unlock(&hugetlb_lock);
724                 return NULL;
725         } else {
726                 h->nr_huge_pages++;
727                 h->surplus_huge_pages++;
728         }
729         spin_unlock(&hugetlb_lock);
730
731         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
732                                         __GFP_REPEAT|__GFP_NOWARN,
733                                         huge_page_order(h));
734
735         if (page && arch_prepare_hugepage(page)) {
736                 __free_pages(page, huge_page_order(h));
737                 return NULL;
738         }
739
740         spin_lock(&hugetlb_lock);
741         if (page) {
742                 /*
743                  * This page is now managed by the hugetlb allocator and has
744                  * no users -- drop the buddy allocator's reference.
745                  */
746                 put_page_testzero(page);
747                 VM_BUG_ON(page_count(page));
748                 nid = page_to_nid(page);
749                 set_compound_page_dtor(page, free_huge_page);
750                 /*
751                  * We incremented the global counters already
752                  */
753                 h->nr_huge_pages_node[nid]++;
754                 h->surplus_huge_pages_node[nid]++;
755                 __count_vm_event(HTLB_BUDDY_PGALLOC);
756         } else {
757                 h->nr_huge_pages--;
758                 h->surplus_huge_pages--;
759                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
760         }
761         spin_unlock(&hugetlb_lock);
762
763         return page;
764 }
765
766 /*
767  * Increase the hugetlb pool such that it can accomodate a reservation
768  * of size 'delta'.
769  */
770 static int gather_surplus_pages(struct hstate *h, int delta)
771 {
772         struct list_head surplus_list;
773         struct page *page, *tmp;
774         int ret, i;
775         int needed, allocated;
776
777         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
778         if (needed <= 0) {
779                 h->resv_huge_pages += delta;
780                 return 0;
781         }
782
783         allocated = 0;
784         INIT_LIST_HEAD(&surplus_list);
785
786         ret = -ENOMEM;
787 retry:
788         spin_unlock(&hugetlb_lock);
789         for (i = 0; i < needed; i++) {
790                 page = alloc_buddy_huge_page(h, NULL, 0);
791                 if (!page) {
792                         /*
793                          * We were not able to allocate enough pages to
794                          * satisfy the entire reservation so we free what
795                          * we've allocated so far.
796                          */
797                         spin_lock(&hugetlb_lock);
798                         needed = 0;
799                         goto free;
800                 }
801
802                 list_add(&page->lru, &surplus_list);
803         }
804         allocated += needed;
805
806         /*
807          * After retaking hugetlb_lock, we need to recalculate 'needed'
808          * because either resv_huge_pages or free_huge_pages may have changed.
809          */
810         spin_lock(&hugetlb_lock);
811         needed = (h->resv_huge_pages + delta) -
812                         (h->free_huge_pages + allocated);
813         if (needed > 0)
814                 goto retry;
815
816         /*
817          * The surplus_list now contains _at_least_ the number of extra pages
818          * needed to accomodate the reservation.  Add the appropriate number
819          * of pages to the hugetlb pool and free the extras back to the buddy
820          * allocator.  Commit the entire reservation here to prevent another
821          * process from stealing the pages as they are added to the pool but
822          * before they are reserved.
823          */
824         needed += allocated;
825         h->resv_huge_pages += delta;
826         ret = 0;
827 free:
828         /* Free the needed pages to the hugetlb pool */
829         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
830                 if ((--needed) < 0)
831                         break;
832                 list_del(&page->lru);
833                 enqueue_huge_page(h, page);
834         }
835
836         /* Free unnecessary surplus pages to the buddy allocator */
837         if (!list_empty(&surplus_list)) {
838                 spin_unlock(&hugetlb_lock);
839                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
840                         list_del(&page->lru);
841                         /*
842                          * The page has a reference count of zero already, so
843                          * call free_huge_page directly instead of using
844                          * put_page.  This must be done with hugetlb_lock
845                          * unlocked which is safe because free_huge_page takes
846                          * hugetlb_lock before deciding how to free the page.
847                          */
848                         free_huge_page(page);
849                 }
850                 spin_lock(&hugetlb_lock);
851         }
852
853         return ret;
854 }
855
856 /*
857  * When releasing a hugetlb pool reservation, any surplus pages that were
858  * allocated to satisfy the reservation must be explicitly freed if they were
859  * never used.
860  */
861 static void return_unused_surplus_pages(struct hstate *h,
862                                         unsigned long unused_resv_pages)
863 {
864         static int nid = -1;
865         struct page *page;
866         unsigned long nr_pages;
867
868         /*
869          * We want to release as many surplus pages as possible, spread
870          * evenly across all nodes. Iterate across all nodes until we
871          * can no longer free unreserved surplus pages. This occurs when
872          * the nodes with surplus pages have no free pages.
873          */
874         unsigned long remaining_iterations = num_online_nodes();
875
876         /* Uncommit the reservation */
877         h->resv_huge_pages -= unused_resv_pages;
878
879         /* Cannot return gigantic pages currently */
880         if (h->order >= MAX_ORDER)
881                 return;
882
883         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
884
885         while (remaining_iterations-- && nr_pages) {
886                 nid = next_node(nid, node_online_map);
887                 if (nid == MAX_NUMNODES)
888                         nid = first_node(node_online_map);
889
890                 if (!h->surplus_huge_pages_node[nid])
891                         continue;
892
893                 if (!list_empty(&h->hugepage_freelists[nid])) {
894                         page = list_entry(h->hugepage_freelists[nid].next,
895                                           struct page, lru);
896                         list_del(&page->lru);
897                         update_and_free_page(h, page);
898                         h->free_huge_pages--;
899                         h->free_huge_pages_node[nid]--;
900                         h->surplus_huge_pages--;
901                         h->surplus_huge_pages_node[nid]--;
902                         nr_pages--;
903                         remaining_iterations = num_online_nodes();
904                 }
905         }
906 }
907
908 /*
909  * Determine if the huge page at addr within the vma has an associated
910  * reservation.  Where it does not we will need to logically increase
911  * reservation and actually increase quota before an allocation can occur.
912  * Where any new reservation would be required the reservation change is
913  * prepared, but not committed.  Once the page has been quota'd allocated
914  * an instantiated the change should be committed via vma_commit_reservation.
915  * No action is required on failure.
916  */
917 static int vma_needs_reservation(struct hstate *h,
918                         struct vm_area_struct *vma, unsigned long addr)
919 {
920         struct address_space *mapping = vma->vm_file->f_mapping;
921         struct inode *inode = mapping->host;
922
923         if (vma->vm_flags & VM_SHARED) {
924                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
925                 return region_chg(&inode->i_mapping->private_list,
926                                                         idx, idx + 1);
927
928         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
929                 return 1;
930
931         } else  {
932                 int err;
933                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
934                 struct resv_map *reservations = vma_resv_map(vma);
935
936                 err = region_chg(&reservations->regions, idx, idx + 1);
937                 if (err < 0)
938                         return err;
939                 return 0;
940         }
941 }
942 static void vma_commit_reservation(struct hstate *h,
943                         struct vm_area_struct *vma, unsigned long addr)
944 {
945         struct address_space *mapping = vma->vm_file->f_mapping;
946         struct inode *inode = mapping->host;
947
948         if (vma->vm_flags & VM_SHARED) {
949                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
950                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
951
952         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
953                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
954                 struct resv_map *reservations = vma_resv_map(vma);
955
956                 /* Mark this page used in the map. */
957                 region_add(&reservations->regions, idx, idx + 1);
958         }
959 }
960
961 static struct page *alloc_huge_page(struct vm_area_struct *vma,
962                                     unsigned long addr, int avoid_reserve)
963 {
964         struct hstate *h = hstate_vma(vma);
965         struct page *page;
966         struct address_space *mapping = vma->vm_file->f_mapping;
967         struct inode *inode = mapping->host;
968         unsigned int chg;
969
970         /*
971          * Processes that did not create the mapping will have no reserves and
972          * will not have accounted against quota. Check that the quota can be
973          * made before satisfying the allocation
974          * MAP_NORESERVE mappings may also need pages and quota allocated
975          * if no reserve mapping overlaps.
976          */
977         chg = vma_needs_reservation(h, vma, addr);
978         if (chg < 0)
979                 return ERR_PTR(chg);
980         if (chg)
981                 if (hugetlb_get_quota(inode->i_mapping, chg))
982                         return ERR_PTR(-ENOSPC);
983
984         spin_lock(&hugetlb_lock);
985         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
986         spin_unlock(&hugetlb_lock);
987
988         if (!page) {
989                 page = alloc_buddy_huge_page(h, vma, addr);
990                 if (!page) {
991                         hugetlb_put_quota(inode->i_mapping, chg);
992                         return ERR_PTR(-VM_FAULT_OOM);
993                 }
994         }
995
996         set_page_refcounted(page);
997         set_page_private(page, (unsigned long) mapping);
998
999         vma_commit_reservation(h, vma, addr);
1000
1001         return page;
1002 }
1003
1004 __attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
1005 {
1006         struct huge_bootmem_page *m;
1007         int nr_nodes = nodes_weight(node_online_map);
1008
1009         while (nr_nodes) {
1010                 void *addr;
1011
1012                 addr = __alloc_bootmem_node_nopanic(
1013                                 NODE_DATA(h->hugetlb_next_nid),
1014                                 huge_page_size(h), huge_page_size(h), 0);
1015
1016                 if (addr) {
1017                         /*
1018                          * Use the beginning of the huge page to store the
1019                          * huge_bootmem_page struct (until gather_bootmem
1020                          * puts them into the mem_map).
1021                          */
1022                         m = addr;
1023                         if (m)
1024                                 goto found;
1025                 }
1026                 hstate_next_node(h);
1027                 nr_nodes--;
1028         }
1029         return 0;
1030
1031 found:
1032         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1033         /* Put them into a private list first because mem_map is not up yet */
1034         list_add(&m->list, &huge_boot_pages);
1035         m->hstate = h;
1036         return 1;
1037 }
1038
1039 static void prep_compound_huge_page(struct page *page, int order)
1040 {
1041         if (unlikely(order > (MAX_ORDER - 1)))
1042                 prep_compound_gigantic_page(page, order);
1043         else
1044                 prep_compound_page(page, order);
1045 }
1046
1047 /* Put bootmem huge pages into the standard lists after mem_map is up */
1048 static void __init gather_bootmem_prealloc(void)
1049 {
1050         struct huge_bootmem_page *m;
1051
1052         list_for_each_entry(m, &huge_boot_pages, list) {
1053                 struct page *page = virt_to_page(m);
1054                 struct hstate *h = m->hstate;
1055                 __ClearPageReserved(page);
1056                 WARN_ON(page_count(page) != 1);
1057                 prep_compound_huge_page(page, h->order);
1058                 prep_new_huge_page(h, page, page_to_nid(page));
1059         }
1060 }
1061
1062 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1063 {
1064         unsigned long i;
1065
1066         for (i = 0; i < h->max_huge_pages; ++i) {
1067                 if (h->order >= MAX_ORDER) {
1068                         if (!alloc_bootmem_huge_page(h))
1069                                 break;
1070                 } else if (!alloc_fresh_huge_page(h))
1071                         break;
1072         }
1073         h->max_huge_pages = i;
1074 }
1075
1076 static void __init hugetlb_init_hstates(void)
1077 {
1078         struct hstate *h;
1079
1080         for_each_hstate(h) {
1081                 /* oversize hugepages were init'ed in early boot */
1082                 if (h->order < MAX_ORDER)
1083                         hugetlb_hstate_alloc_pages(h);
1084         }
1085 }
1086
1087 static char * __init memfmt(char *buf, unsigned long n)
1088 {
1089         if (n >= (1UL << 30))
1090                 sprintf(buf, "%lu GB", n >> 30);
1091         else if (n >= (1UL << 20))
1092                 sprintf(buf, "%lu MB", n >> 20);
1093         else
1094                 sprintf(buf, "%lu KB", n >> 10);
1095         return buf;
1096 }
1097
1098 static void __init report_hugepages(void)
1099 {
1100         struct hstate *h;
1101
1102         for_each_hstate(h) {
1103                 char buf[32];
1104                 printk(KERN_INFO "HugeTLB registered %s page size, "
1105                                  "pre-allocated %ld pages\n",
1106                         memfmt(buf, huge_page_size(h)),
1107                         h->free_huge_pages);
1108         }
1109 }
1110
1111 #ifdef CONFIG_HIGHMEM
1112 static void try_to_free_low(struct hstate *h, unsigned long count)
1113 {
1114         int i;
1115
1116         if (h->order >= MAX_ORDER)
1117                 return;
1118
1119         for (i = 0; i < MAX_NUMNODES; ++i) {
1120                 struct page *page, *next;
1121                 struct list_head *freel = &h->hugepage_freelists[i];
1122                 list_for_each_entry_safe(page, next, freel, lru) {
1123                         if (count >= h->nr_huge_pages)
1124                                 return;
1125                         if (PageHighMem(page))
1126                                 continue;
1127                         list_del(&page->lru);
1128                         update_and_free_page(h, page);
1129                         h->free_huge_pages--;
1130                         h->free_huge_pages_node[page_to_nid(page)]--;
1131                 }
1132         }
1133 }
1134 #else
1135 static inline void try_to_free_low(struct hstate *h, unsigned long count)
1136 {
1137 }
1138 #endif
1139
1140 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1141 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
1142 {
1143         unsigned long min_count, ret;
1144
1145         if (h->order >= MAX_ORDER)
1146                 return h->max_huge_pages;
1147
1148         /*
1149          * Increase the pool size
1150          * First take pages out of surplus state.  Then make up the
1151          * remaining difference by allocating fresh huge pages.
1152          *
1153          * We might race with alloc_buddy_huge_page() here and be unable
1154          * to convert a surplus huge page to a normal huge page. That is
1155          * not critical, though, it just means the overall size of the
1156          * pool might be one hugepage larger than it needs to be, but
1157          * within all the constraints specified by the sysctls.
1158          */
1159         spin_lock(&hugetlb_lock);
1160         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1161                 if (!adjust_pool_surplus(h, -1))
1162                         break;
1163         }
1164
1165         while (count > persistent_huge_pages(h)) {
1166                 /*
1167                  * If this allocation races such that we no longer need the
1168                  * page, free_huge_page will handle it by freeing the page
1169                  * and reducing the surplus.
1170                  */
1171                 spin_unlock(&hugetlb_lock);
1172                 ret = alloc_fresh_huge_page(h);
1173                 spin_lock(&hugetlb_lock);
1174                 if (!ret)
1175                         goto out;
1176
1177         }
1178
1179         /*
1180          * Decrease the pool size
1181          * First return free pages to the buddy allocator (being careful
1182          * to keep enough around to satisfy reservations).  Then place
1183          * pages into surplus state as needed so the pool will shrink
1184          * to the desired size as pages become free.
1185          *
1186          * By placing pages into the surplus state independent of the
1187          * overcommit value, we are allowing the surplus pool size to
1188          * exceed overcommit. There are few sane options here. Since
1189          * alloc_buddy_huge_page() is checking the global counter,
1190          * though, we'll note that we're not allowed to exceed surplus
1191          * and won't grow the pool anywhere else. Not until one of the
1192          * sysctls are changed, or the surplus pages go out of use.
1193          */
1194         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1195         min_count = max(count, min_count);
1196         try_to_free_low(h, min_count);
1197         while (min_count < persistent_huge_pages(h)) {
1198                 struct page *page = dequeue_huge_page(h);
1199                 if (!page)
1200                         break;
1201                 update_and_free_page(h, page);
1202         }
1203         while (count < persistent_huge_pages(h)) {
1204                 if (!adjust_pool_surplus(h, 1))
1205                         break;
1206         }
1207 out:
1208         ret = persistent_huge_pages(h);
1209         spin_unlock(&hugetlb_lock);
1210         return ret;
1211 }
1212
1213 #define HSTATE_ATTR_RO(_name) \
1214         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1215
1216 #define HSTATE_ATTR(_name) \
1217         static struct kobj_attribute _name##_attr = \
1218                 __ATTR(_name, 0644, _name##_show, _name##_store)
1219
1220 static struct kobject *hugepages_kobj;
1221 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1222
1223 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1224 {
1225         int i;
1226         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1227                 if (hstate_kobjs[i] == kobj)
1228                         return &hstates[i];
1229         BUG();
1230         return NULL;
1231 }
1232
1233 static ssize_t nr_hugepages_show(struct kobject *kobj,
1234                                         struct kobj_attribute *attr, char *buf)
1235 {
1236         struct hstate *h = kobj_to_hstate(kobj);
1237         return sprintf(buf, "%lu\n", h->nr_huge_pages);
1238 }
1239 static ssize_t nr_hugepages_store(struct kobject *kobj,
1240                 struct kobj_attribute *attr, const char *buf, size_t count)
1241 {
1242         int err;
1243         unsigned long input;
1244         struct hstate *h = kobj_to_hstate(kobj);
1245
1246         err = strict_strtoul(buf, 10, &input);
1247         if (err)
1248                 return 0;
1249
1250         h->max_huge_pages = set_max_huge_pages(h, input);
1251
1252         return count;
1253 }
1254 HSTATE_ATTR(nr_hugepages);
1255
1256 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1257                                         struct kobj_attribute *attr, char *buf)
1258 {
1259         struct hstate *h = kobj_to_hstate(kobj);
1260         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1261 }
1262 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1263                 struct kobj_attribute *attr, const char *buf, size_t count)
1264 {
1265         int err;
1266         unsigned long input;
1267         struct hstate *h = kobj_to_hstate(kobj);
1268
1269         err = strict_strtoul(buf, 10, &input);
1270         if (err)
1271                 return 0;
1272
1273         spin_lock(&hugetlb_lock);
1274         h->nr_overcommit_huge_pages = input;
1275         spin_unlock(&hugetlb_lock);
1276
1277         return count;
1278 }
1279 HSTATE_ATTR(nr_overcommit_hugepages);
1280
1281 static ssize_t free_hugepages_show(struct kobject *kobj,
1282                                         struct kobj_attribute *attr, char *buf)
1283 {
1284         struct hstate *h = kobj_to_hstate(kobj);
1285         return sprintf(buf, "%lu\n", h->free_huge_pages);
1286 }
1287 HSTATE_ATTR_RO(free_hugepages);
1288
1289 static ssize_t resv_hugepages_show(struct kobject *kobj,
1290                                         struct kobj_attribute *attr, char *buf)
1291 {
1292         struct hstate *h = kobj_to_hstate(kobj);
1293         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1294 }
1295 HSTATE_ATTR_RO(resv_hugepages);
1296
1297 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1298                                         struct kobj_attribute *attr, char *buf)
1299 {
1300         struct hstate *h = kobj_to_hstate(kobj);
1301         return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1302 }
1303 HSTATE_ATTR_RO(surplus_hugepages);
1304
1305 static struct attribute *hstate_attrs[] = {
1306         &nr_hugepages_attr.attr,
1307         &nr_overcommit_hugepages_attr.attr,
1308         &free_hugepages_attr.attr,
1309         &resv_hugepages_attr.attr,
1310         &surplus_hugepages_attr.attr,
1311         NULL,
1312 };
1313
1314 static struct attribute_group hstate_attr_group = {
1315         .attrs = hstate_attrs,
1316 };
1317
1318 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1319 {
1320         int retval;
1321
1322         hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1323                                                         hugepages_kobj);
1324         if (!hstate_kobjs[h - hstates])
1325                 return -ENOMEM;
1326
1327         retval = sysfs_create_group(hstate_kobjs[h - hstates],
1328                                                         &hstate_attr_group);
1329         if (retval)
1330                 kobject_put(hstate_kobjs[h - hstates]);
1331
1332         return retval;
1333 }
1334
1335 static void __init hugetlb_sysfs_init(void)
1336 {
1337         struct hstate *h;
1338         int err;
1339
1340         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1341         if (!hugepages_kobj)
1342                 return;
1343
1344         for_each_hstate(h) {
1345                 err = hugetlb_sysfs_add_hstate(h);
1346                 if (err)
1347                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1348                                                                 h->name);
1349         }
1350 }
1351
1352 static void __exit hugetlb_exit(void)
1353 {
1354         struct hstate *h;
1355
1356         for_each_hstate(h) {
1357                 kobject_put(hstate_kobjs[h - hstates]);
1358         }
1359
1360         kobject_put(hugepages_kobj);
1361 }
1362 module_exit(hugetlb_exit);
1363
1364 static int __init hugetlb_init(void)
1365 {
1366         /* Some platform decide whether they support huge pages at boot
1367          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1368          * there is no such support
1369          */
1370         if (HPAGE_SHIFT == 0)
1371                 return 0;
1372
1373         if (!size_to_hstate(default_hstate_size)) {
1374                 default_hstate_size = HPAGE_SIZE;
1375                 if (!size_to_hstate(default_hstate_size))
1376                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1377         }
1378         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1379         if (default_hstate_max_huge_pages)
1380                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1381
1382         hugetlb_init_hstates();
1383
1384         gather_bootmem_prealloc();
1385
1386         report_hugepages();
1387
1388         hugetlb_sysfs_init();
1389
1390         return 0;
1391 }
1392 module_init(hugetlb_init);
1393
1394 /* Should be called on processing a hugepagesz=... option */
1395 void __init hugetlb_add_hstate(unsigned order)
1396 {
1397         struct hstate *h;
1398         unsigned long i;
1399
1400         if (size_to_hstate(PAGE_SIZE << order)) {
1401                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1402                 return;
1403         }
1404         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1405         BUG_ON(order == 0);
1406         h = &hstates[max_hstate++];
1407         h->order = order;
1408         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1409         h->nr_huge_pages = 0;
1410         h->free_huge_pages = 0;
1411         for (i = 0; i < MAX_NUMNODES; ++i)
1412                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1413         h->hugetlb_next_nid = first_node(node_online_map);
1414         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1415                                         huge_page_size(h)/1024);
1416
1417         parsed_hstate = h;
1418 }
1419
1420 static int __init hugetlb_nrpages_setup(char *s)
1421 {
1422         unsigned long *mhp;
1423         static unsigned long *last_mhp;
1424
1425         /*
1426          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1427          * so this hugepages= parameter goes to the "default hstate".
1428          */
1429         if (!max_hstate)
1430                 mhp = &default_hstate_max_huge_pages;
1431         else
1432                 mhp = &parsed_hstate->max_huge_pages;
1433
1434         if (mhp == last_mhp) {
1435                 printk(KERN_WARNING "hugepages= specified twice without "
1436                         "interleaving hugepagesz=, ignoring\n");
1437                 return 1;
1438         }
1439
1440         if (sscanf(s, "%lu", mhp) <= 0)
1441                 *mhp = 0;
1442
1443         /*
1444          * Global state is always initialized later in hugetlb_init.
1445          * But we need to allocate >= MAX_ORDER hstates here early to still
1446          * use the bootmem allocator.
1447          */
1448         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1449                 hugetlb_hstate_alloc_pages(parsed_hstate);
1450
1451         last_mhp = mhp;
1452
1453         return 1;
1454 }
1455 __setup("hugepages=", hugetlb_nrpages_setup);
1456
1457 static int __init hugetlb_default_setup(char *s)
1458 {
1459         default_hstate_size = memparse(s, &s);
1460         return 1;
1461 }
1462 __setup("default_hugepagesz=", hugetlb_default_setup);
1463
1464 static unsigned int cpuset_mems_nr(unsigned int *array)
1465 {
1466         int node;
1467         unsigned int nr = 0;
1468
1469         for_each_node_mask(node, cpuset_current_mems_allowed)
1470                 nr += array[node];
1471
1472         return nr;
1473 }
1474
1475 #ifdef CONFIG_SYSCTL
1476 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1477                            struct file *file, void __user *buffer,
1478                            size_t *length, loff_t *ppos)
1479 {
1480         struct hstate *h = &default_hstate;
1481         unsigned long tmp;
1482
1483         if (!write)
1484                 tmp = h->max_huge_pages;
1485
1486         table->data = &tmp;
1487         table->maxlen = sizeof(unsigned long);
1488         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1489
1490         if (write)
1491                 h->max_huge_pages = set_max_huge_pages(h, tmp);
1492
1493         return 0;
1494 }
1495
1496 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1497                         struct file *file, void __user *buffer,
1498                         size_t *length, loff_t *ppos)
1499 {
1500         proc_dointvec(table, write, file, buffer, length, ppos);
1501         if (hugepages_treat_as_movable)
1502                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1503         else
1504                 htlb_alloc_mask = GFP_HIGHUSER;
1505         return 0;
1506 }
1507
1508 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1509                         struct file *file, void __user *buffer,
1510                         size_t *length, loff_t *ppos)
1511 {
1512         struct hstate *h = &default_hstate;
1513         unsigned long tmp;
1514
1515         if (!write)
1516                 tmp = h->nr_overcommit_huge_pages;
1517
1518         table->data = &tmp;
1519         table->maxlen = sizeof(unsigned long);
1520         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1521
1522         if (write) {
1523                 spin_lock(&hugetlb_lock);
1524                 h->nr_overcommit_huge_pages = tmp;
1525                 spin_unlock(&hugetlb_lock);
1526         }
1527
1528         return 0;
1529 }
1530
1531 #endif /* CONFIG_SYSCTL */
1532
1533 void hugetlb_report_meminfo(struct seq_file *m)
1534 {
1535         struct hstate *h = &default_hstate;
1536         seq_printf(m,
1537                         "HugePages_Total:   %5lu\n"
1538                         "HugePages_Free:    %5lu\n"
1539                         "HugePages_Rsvd:    %5lu\n"
1540                         "HugePages_Surp:    %5lu\n"
1541                         "Hugepagesize:   %8lu kB\n",
1542                         h->nr_huge_pages,
1543                         h->free_huge_pages,
1544                         h->resv_huge_pages,
1545                         h->surplus_huge_pages,
1546                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1547 }
1548
1549 int hugetlb_report_node_meminfo(int nid, char *buf)
1550 {
1551         struct hstate *h = &default_hstate;
1552         return sprintf(buf,
1553                 "Node %d HugePages_Total: %5u\n"
1554                 "Node %d HugePages_Free:  %5u\n"
1555                 "Node %d HugePages_Surp:  %5u\n",
1556                 nid, h->nr_huge_pages_node[nid],
1557                 nid, h->free_huge_pages_node[nid],
1558                 nid, h->surplus_huge_pages_node[nid]);
1559 }
1560
1561 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1562 unsigned long hugetlb_total_pages(void)
1563 {
1564         struct hstate *h = &default_hstate;
1565         return h->nr_huge_pages * pages_per_huge_page(h);
1566 }
1567
1568 static int hugetlb_acct_memory(struct hstate *h, long delta)
1569 {
1570         int ret = -ENOMEM;
1571
1572         spin_lock(&hugetlb_lock);
1573         /*
1574          * When cpuset is configured, it breaks the strict hugetlb page
1575          * reservation as the accounting is done on a global variable. Such
1576          * reservation is completely rubbish in the presence of cpuset because
1577          * the reservation is not checked against page availability for the
1578          * current cpuset. Application can still potentially OOM'ed by kernel
1579          * with lack of free htlb page in cpuset that the task is in.
1580          * Attempt to enforce strict accounting with cpuset is almost
1581          * impossible (or too ugly) because cpuset is too fluid that
1582          * task or memory node can be dynamically moved between cpusets.
1583          *
1584          * The change of semantics for shared hugetlb mapping with cpuset is
1585          * undesirable. However, in order to preserve some of the semantics,
1586          * we fall back to check against current free page availability as
1587          * a best attempt and hopefully to minimize the impact of changing
1588          * semantics that cpuset has.
1589          */
1590         if (delta > 0) {
1591                 if (gather_surplus_pages(h, delta) < 0)
1592                         goto out;
1593
1594                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1595                         return_unused_surplus_pages(h, delta);
1596                         goto out;
1597                 }
1598         }
1599
1600         ret = 0;
1601         if (delta < 0)
1602                 return_unused_surplus_pages(h, (unsigned long) -delta);
1603
1604 out:
1605         spin_unlock(&hugetlb_lock);
1606         return ret;
1607 }
1608
1609 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1610 {
1611         struct resv_map *reservations = vma_resv_map(vma);
1612
1613         /*
1614          * This new VMA should share its siblings reservation map if present.
1615          * The VMA will only ever have a valid reservation map pointer where
1616          * it is being copied for another still existing VMA.  As that VMA
1617          * has a reference to the reservation map it cannot dissappear until
1618          * after this open call completes.  It is therefore safe to take a
1619          * new reference here without additional locking.
1620          */
1621         if (reservations)
1622                 kref_get(&reservations->refs);
1623 }
1624
1625 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1626 {
1627         struct hstate *h = hstate_vma(vma);
1628         struct resv_map *reservations = vma_resv_map(vma);
1629         unsigned long reserve;
1630         unsigned long start;
1631         unsigned long end;
1632
1633         if (reservations) {
1634                 start = vma_hugecache_offset(h, vma, vma->vm_start);
1635                 end = vma_hugecache_offset(h, vma, vma->vm_end);
1636
1637                 reserve = (end - start) -
1638                         region_count(&reservations->regions, start, end);
1639
1640                 kref_put(&reservations->refs, resv_map_release);
1641
1642                 if (reserve) {
1643                         hugetlb_acct_memory(h, -reserve);
1644                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1645                 }
1646         }
1647 }
1648
1649 /*
1650  * We cannot handle pagefaults against hugetlb pages at all.  They cause
1651  * handle_mm_fault() to try to instantiate regular-sized pages in the
1652  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
1653  * this far.
1654  */
1655 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1656 {
1657         BUG();
1658         return 0;
1659 }
1660
1661 struct vm_operations_struct hugetlb_vm_ops = {
1662         .fault = hugetlb_vm_op_fault,
1663         .open = hugetlb_vm_op_open,
1664         .close = hugetlb_vm_op_close,
1665 };
1666
1667 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1668                                 int writable)
1669 {
1670         pte_t entry;
1671
1672         if (writable) {
1673                 entry =
1674                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1675         } else {
1676                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1677         }
1678         entry = pte_mkyoung(entry);
1679         entry = pte_mkhuge(entry);
1680
1681         return entry;
1682 }
1683
1684 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1685                                    unsigned long address, pte_t *ptep)
1686 {
1687         pte_t entry;
1688
1689         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1690         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1691                 update_mmu_cache(vma, address, entry);
1692         }
1693 }
1694
1695
1696 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1697                             struct vm_area_struct *vma)
1698 {
1699         pte_t *src_pte, *dst_pte, entry;
1700         struct page *ptepage;
1701         unsigned long addr;
1702         int cow;
1703         struct hstate *h = hstate_vma(vma);
1704         unsigned long sz = huge_page_size(h);
1705
1706         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1707
1708         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1709                 src_pte = huge_pte_offset(src, addr);
1710                 if (!src_pte)
1711                         continue;
1712                 dst_pte = huge_pte_alloc(dst, addr, sz);
1713                 if (!dst_pte)
1714                         goto nomem;
1715
1716                 /* If the pagetables are shared don't copy or take references */
1717                 if (dst_pte == src_pte)
1718                         continue;
1719
1720                 spin_lock(&dst->page_table_lock);
1721                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1722                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1723                         if (cow)
1724                                 huge_ptep_set_wrprotect(src, addr, src_pte);
1725                         entry = huge_ptep_get(src_pte);
1726                         ptepage = pte_page(entry);
1727                         get_page(ptepage);
1728                         set_huge_pte_at(dst, addr, dst_pte, entry);
1729                 }
1730                 spin_unlock(&src->page_table_lock);
1731                 spin_unlock(&dst->page_table_lock);
1732         }
1733         return 0;
1734
1735 nomem:
1736         return -ENOMEM;
1737 }
1738
1739 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1740                             unsigned long end, struct page *ref_page)
1741 {
1742         struct mm_struct *mm = vma->vm_mm;
1743         unsigned long address;
1744         pte_t *ptep;
1745         pte_t pte;
1746         struct page *page;
1747         struct page *tmp;
1748         struct hstate *h = hstate_vma(vma);
1749         unsigned long sz = huge_page_size(h);
1750
1751         /*
1752          * A page gathering list, protected by per file i_mmap_lock. The
1753          * lock is used to avoid list corruption from multiple unmapping
1754          * of the same page since we are using page->lru.
1755          */
1756         LIST_HEAD(page_list);
1757
1758         WARN_ON(!is_vm_hugetlb_page(vma));
1759         BUG_ON(start & ~huge_page_mask(h));
1760         BUG_ON(end & ~huge_page_mask(h));
1761
1762         mmu_notifier_invalidate_range_start(mm, start, end);
1763         spin_lock(&mm->page_table_lock);
1764         for (address = start; address < end; address += sz) {
1765                 ptep = huge_pte_offset(mm, address);
1766                 if (!ptep)
1767                         continue;
1768
1769                 if (huge_pmd_unshare(mm, &address, ptep))
1770                         continue;
1771
1772                 /*
1773                  * If a reference page is supplied, it is because a specific
1774                  * page is being unmapped, not a range. Ensure the page we
1775                  * are about to unmap is the actual page of interest.
1776                  */
1777                 if (ref_page) {
1778                         pte = huge_ptep_get(ptep);
1779                         if (huge_pte_none(pte))
1780                                 continue;
1781                         page = pte_page(pte);
1782                         if (page != ref_page)
1783                                 continue;
1784
1785                         /*
1786                          * Mark the VMA as having unmapped its page so that
1787                          * future faults in this VMA will fail rather than
1788                          * looking like data was lost
1789                          */
1790                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1791                 }
1792
1793                 pte = huge_ptep_get_and_clear(mm, address, ptep);
1794                 if (huge_pte_none(pte))
1795                         continue;
1796
1797                 page = pte_page(pte);
1798                 if (pte_dirty(pte))
1799                         set_page_dirty(page);
1800                 list_add(&page->lru, &page_list);
1801         }
1802         spin_unlock(&mm->page_table_lock);
1803         flush_tlb_range(vma, start, end);
1804         mmu_notifier_invalidate_range_end(mm, start, end);
1805         list_for_each_entry_safe(page, tmp, &page_list, lru) {
1806                 list_del(&page->lru);
1807                 put_page(page);
1808         }
1809 }
1810
1811 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1812                           unsigned long end, struct page *ref_page)
1813 {
1814         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1815         __unmap_hugepage_range(vma, start, end, ref_page);
1816         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1817 }
1818
1819 /*
1820  * This is called when the original mapper is failing to COW a MAP_PRIVATE
1821  * mappping it owns the reserve page for. The intention is to unmap the page
1822  * from other VMAs and let the children be SIGKILLed if they are faulting the
1823  * same region.
1824  */
1825 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1826                                 struct page *page, unsigned long address)
1827 {
1828         struct hstate *h = hstate_vma(vma);
1829         struct vm_area_struct *iter_vma;
1830         struct address_space *mapping;
1831         struct prio_tree_iter iter;
1832         pgoff_t pgoff;
1833
1834         /*
1835          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1836          * from page cache lookup which is in HPAGE_SIZE units.
1837          */
1838         address = address & huge_page_mask(h);
1839         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1840                 + (vma->vm_pgoff >> PAGE_SHIFT);
1841         mapping = (struct address_space *)page_private(page);
1842
1843         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1844                 /* Do not unmap the current VMA */
1845                 if (iter_vma == vma)
1846                         continue;
1847
1848                 /*
1849                  * Unmap the page from other VMAs without their own reserves.
1850                  * They get marked to be SIGKILLed if they fault in these
1851                  * areas. This is because a future no-page fault on this VMA
1852                  * could insert a zeroed page instead of the data existing
1853                  * from the time of fork. This would look like data corruption
1854                  */
1855                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1856                         unmap_hugepage_range(iter_vma,
1857                                 address, address + huge_page_size(h),
1858                                 page);
1859         }
1860
1861         return 1;
1862 }
1863
1864 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1865                         unsigned long address, pte_t *ptep, pte_t pte,
1866                         struct page *pagecache_page)
1867 {
1868         struct hstate *h = hstate_vma(vma);
1869         struct page *old_page, *new_page;
1870         int avoidcopy;
1871         int outside_reserve = 0;
1872
1873         old_page = pte_page(pte);
1874
1875 retry_avoidcopy:
1876         /* If no-one else is actually using this page, avoid the copy
1877          * and just make the page writable */
1878         avoidcopy = (page_count(old_page) == 1);
1879         if (avoidcopy) {
1880                 set_huge_ptep_writable(vma, address, ptep);
1881                 return 0;
1882         }
1883
1884         /*
1885          * If the process that created a MAP_PRIVATE mapping is about to
1886          * perform a COW due to a shared page count, attempt to satisfy
1887          * the allocation without using the existing reserves. The pagecache
1888          * page is used to determine if the reserve at this address was
1889          * consumed or not. If reserves were used, a partial faulted mapping
1890          * at the time of fork() could consume its reserves on COW instead
1891          * of the full address range.
1892          */
1893         if (!(vma->vm_flags & VM_SHARED) &&
1894                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1895                         old_page != pagecache_page)
1896                 outside_reserve = 1;
1897
1898         page_cache_get(old_page);
1899         new_page = alloc_huge_page(vma, address, outside_reserve);
1900
1901         if (IS_ERR(new_page)) {
1902                 page_cache_release(old_page);
1903
1904                 /*
1905                  * If a process owning a MAP_PRIVATE mapping fails to COW,
1906                  * it is due to references held by a child and an insufficient
1907                  * huge page pool. To guarantee the original mappers
1908                  * reliability, unmap the page from child processes. The child
1909                  * may get SIGKILLed if it later faults.
1910                  */
1911                 if (outside_reserve) {
1912                         BUG_ON(huge_pte_none(pte));
1913                         if (unmap_ref_private(mm, vma, old_page, address)) {
1914                                 BUG_ON(page_count(old_page) != 1);
1915                                 BUG_ON(huge_pte_none(pte));
1916                                 goto retry_avoidcopy;
1917                         }
1918                         WARN_ON_ONCE(1);
1919                 }
1920
1921                 return -PTR_ERR(new_page);
1922         }
1923
1924         spin_unlock(&mm->page_table_lock);
1925         copy_huge_page(new_page, old_page, address, vma);
1926         __SetPageUptodate(new_page);
1927         spin_lock(&mm->page_table_lock);
1928
1929         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1930         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1931                 /* Break COW */
1932                 huge_ptep_clear_flush(vma, address, ptep);
1933                 set_huge_pte_at(mm, address, ptep,
1934                                 make_huge_pte(vma, new_page, 1));
1935                 /* Make the old page be freed below */
1936                 new_page = old_page;
1937         }
1938         page_cache_release(new_page);
1939         page_cache_release(old_page);
1940         return 0;
1941 }
1942
1943 /* Return the pagecache page at a given address within a VMA */
1944 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1945                         struct vm_area_struct *vma, unsigned long address)
1946 {
1947         struct address_space *mapping;
1948         pgoff_t idx;
1949
1950         mapping = vma->vm_file->f_mapping;
1951         idx = vma_hugecache_offset(h, vma, address);
1952
1953         return find_lock_page(mapping, idx);
1954 }
1955
1956 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1957                         unsigned long address, pte_t *ptep, int write_access)
1958 {
1959         struct hstate *h = hstate_vma(vma);
1960         int ret = VM_FAULT_SIGBUS;
1961         pgoff_t idx;
1962         unsigned long size;
1963         struct page *page;
1964         struct address_space *mapping;
1965         pte_t new_pte;
1966
1967         /*
1968          * Currently, we are forced to kill the process in the event the
1969          * original mapper has unmapped pages from the child due to a failed
1970          * COW. Warn that such a situation has occured as it may not be obvious
1971          */
1972         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1973                 printk(KERN_WARNING
1974                         "PID %d killed due to inadequate hugepage pool\n",
1975                         current->pid);
1976                 return ret;
1977         }
1978
1979         mapping = vma->vm_file->f_mapping;
1980         idx = vma_hugecache_offset(h, vma, address);
1981
1982         /*
1983          * Use page lock to guard against racing truncation
1984          * before we get page_table_lock.
1985          */
1986 retry:
1987         page = find_lock_page(mapping, idx);
1988         if (!page) {
1989                 size = i_size_read(mapping->host) >> huge_page_shift(h);
1990                 if (idx >= size)
1991                         goto out;
1992                 page = alloc_huge_page(vma, address, 0);
1993                 if (IS_ERR(page)) {
1994                         ret = -PTR_ERR(page);
1995                         goto out;
1996                 }
1997                 clear_huge_page(page, address, huge_page_size(h));
1998                 __SetPageUptodate(page);
1999
2000                 if (vma->vm_flags & VM_SHARED) {
2001                         int err;
2002                         struct inode *inode = mapping->host;
2003
2004                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2005                         if (err) {
2006                                 put_page(page);
2007                                 if (err == -EEXIST)
2008                                         goto retry;
2009                                 goto out;
2010                         }
2011
2012                         spin_lock(&inode->i_lock);
2013                         inode->i_blocks += blocks_per_huge_page(h);
2014                         spin_unlock(&inode->i_lock);
2015                 } else
2016                         lock_page(page);
2017         }
2018
2019         /*
2020          * If we are going to COW a private mapping later, we examine the
2021          * pending reservations for this page now. This will ensure that
2022          * any allocations necessary to record that reservation occur outside
2023          * the spinlock.
2024          */
2025         if (write_access && !(vma->vm_flags & VM_SHARED))
2026                 if (vma_needs_reservation(h, vma, address) < 0) {
2027                         ret = VM_FAULT_OOM;
2028                         goto backout_unlocked;
2029                 }
2030
2031         spin_lock(&mm->page_table_lock);
2032         size = i_size_read(mapping->host) >> huge_page_shift(h);
2033         if (idx >= size)
2034                 goto backout;
2035
2036         ret = 0;
2037         if (!huge_pte_none(huge_ptep_get(ptep)))
2038                 goto backout;
2039
2040         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2041                                 && (vma->vm_flags & VM_SHARED)));
2042         set_huge_pte_at(mm, address, ptep, new_pte);
2043
2044         if (write_access && !(vma->vm_flags & VM_SHARED)) {
2045                 /* Optimization, do the COW without a second fault */
2046                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2047         }
2048
2049         spin_unlock(&mm->page_table_lock);
2050         unlock_page(page);
2051 out:
2052         return ret;
2053
2054 backout:
2055         spin_unlock(&mm->page_table_lock);
2056 backout_unlocked:
2057         unlock_page(page);
2058         put_page(page);
2059         goto out;
2060 }
2061
2062 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2063                         unsigned long address, int write_access)
2064 {
2065         pte_t *ptep;
2066         pte_t entry;
2067         int ret;
2068         struct page *pagecache_page = NULL;
2069         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2070         struct hstate *h = hstate_vma(vma);
2071
2072         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2073         if (!ptep)
2074                 return VM_FAULT_OOM;
2075
2076         /*
2077          * Serialize hugepage allocation and instantiation, so that we don't
2078          * get spurious allocation failures if two CPUs race to instantiate
2079          * the same page in the page cache.
2080          */
2081         mutex_lock(&hugetlb_instantiation_mutex);
2082         entry = huge_ptep_get(ptep);
2083         if (huge_pte_none(entry)) {
2084                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
2085                 goto out_mutex;
2086         }
2087
2088         ret = 0;
2089
2090         /*
2091          * If we are going to COW the mapping later, we examine the pending
2092          * reservations for this page now. This will ensure that any
2093          * allocations necessary to record that reservation occur outside the
2094          * spinlock. For private mappings, we also lookup the pagecache
2095          * page now as it is used to determine if a reservation has been
2096          * consumed.
2097          */
2098         if (write_access && !pte_write(entry)) {
2099                 if (vma_needs_reservation(h, vma, address) < 0) {
2100                         ret = VM_FAULT_OOM;
2101                         goto out_mutex;
2102                 }
2103
2104                 if (!(vma->vm_flags & VM_SHARED))
2105                         pagecache_page = hugetlbfs_pagecache_page(h,
2106                                                                 vma, address);
2107         }
2108
2109         spin_lock(&mm->page_table_lock);
2110         /* Check for a racing update before calling hugetlb_cow */
2111         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2112                 goto out_page_table_lock;
2113
2114
2115         if (write_access) {
2116                 if (!pte_write(entry)) {
2117                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2118                                                         pagecache_page);
2119                         goto out_page_table_lock;
2120                 }
2121                 entry = pte_mkdirty(entry);
2122         }
2123         entry = pte_mkyoung(entry);
2124         if (huge_ptep_set_access_flags(vma, address, ptep, entry, write_access))
2125                 update_mmu_cache(vma, address, entry);
2126
2127 out_page_table_lock:
2128         spin_unlock(&mm->page_table_lock);
2129
2130         if (pagecache_page) {
2131                 unlock_page(pagecache_page);
2132                 put_page(pagecache_page);
2133         }
2134
2135 out_mutex:
2136         mutex_unlock(&hugetlb_instantiation_mutex);
2137
2138         return ret;
2139 }
2140
2141 /* Can be overriden by architectures */
2142 __attribute__((weak)) struct page *
2143 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2144                pud_t *pud, int write)
2145 {
2146         BUG();
2147         return NULL;
2148 }
2149
2150 static int huge_zeropage_ok(pte_t *ptep, int write, int shared)
2151 {
2152         if (!ptep || write || shared)
2153                 return 0;
2154         else
2155                 return huge_pte_none(huge_ptep_get(ptep));
2156 }
2157
2158 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2159                         struct page **pages, struct vm_area_struct **vmas,
2160                         unsigned long *position, int *length, int i,
2161                         int write)
2162 {
2163         unsigned long pfn_offset;
2164         unsigned long vaddr = *position;
2165         int remainder = *length;
2166         struct hstate *h = hstate_vma(vma);
2167         int zeropage_ok = 0;
2168         int shared = vma->vm_flags & VM_SHARED;
2169
2170         spin_lock(&mm->page_table_lock);
2171         while (vaddr < vma->vm_end && remainder) {
2172                 pte_t *pte;
2173                 struct page *page;
2174
2175                 /*
2176                  * Some archs (sparc64, sh*) have multiple pte_ts to
2177                  * each hugepage.  We have to make * sure we get the
2178                  * first, for the page indexing below to work.
2179                  */
2180                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2181                 if (huge_zeropage_ok(pte, write, shared))
2182                         zeropage_ok = 1;
2183
2184                 if (!pte ||
2185                     (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) ||
2186                     (write && !pte_write(huge_ptep_get(pte)))) {
2187                         int ret;
2188
2189                         spin_unlock(&mm->page_table_lock);
2190                         ret = hugetlb_fault(mm, vma, vaddr, write);
2191                         spin_lock(&mm->page_table_lock);
2192                         if (!(ret & VM_FAULT_ERROR))
2193                                 continue;
2194
2195                         remainder = 0;
2196                         if (!i)
2197                                 i = -EFAULT;
2198                         break;
2199                 }
2200
2201                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2202                 page = pte_page(huge_ptep_get(pte));
2203 same_page:
2204                 if (pages) {
2205                         if (zeropage_ok)
2206                                 pages[i] = ZERO_PAGE(0);
2207                         else
2208                                 pages[i] = mem_map_offset(page, pfn_offset);
2209                         get_page(pages[i]);
2210                 }
2211
2212                 if (vmas)
2213                         vmas[i] = vma;
2214
2215                 vaddr += PAGE_SIZE;
2216                 ++pfn_offset;
2217                 --remainder;
2218                 ++i;
2219                 if (vaddr < vma->vm_end && remainder &&
2220                                 pfn_offset < pages_per_huge_page(h)) {
2221                         /*
2222                          * We use pfn_offset to avoid touching the pageframes
2223                          * of this compound page.
2224                          */
2225                         goto same_page;
2226                 }
2227         }
2228         spin_unlock(&mm->page_table_lock);
2229         *length = remainder;
2230         *position = vaddr;
2231
2232         return i;
2233 }
2234
2235 void hugetlb_change_protection(struct vm_area_struct *vma,
2236                 unsigned long address, unsigned long end, pgprot_t newprot)
2237 {
2238         struct mm_struct *mm = vma->vm_mm;
2239         unsigned long start = address;
2240         pte_t *ptep;
2241         pte_t pte;
2242         struct hstate *h = hstate_vma(vma);
2243
2244         BUG_ON(address >= end);
2245         flush_cache_range(vma, address, end);
2246
2247         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2248         spin_lock(&mm->page_table_lock);
2249         for (; address < end; address += huge_page_size(h)) {
2250                 ptep = huge_pte_offset(mm, address);
2251                 if (!ptep)
2252                         continue;
2253                 if (huge_pmd_unshare(mm, &address, ptep))
2254                         continue;
2255                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2256                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2257                         pte = pte_mkhuge(pte_modify(pte, newprot));
2258                         set_huge_pte_at(mm, address, ptep, pte);
2259                 }
2260         }
2261         spin_unlock(&mm->page_table_lock);
2262         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2263
2264         flush_tlb_range(vma, start, end);
2265 }
2266
2267 int hugetlb_reserve_pages(struct inode *inode,
2268                                         long from, long to,
2269                                         struct vm_area_struct *vma)
2270 {
2271         long ret, chg;
2272         struct hstate *h = hstate_inode(inode);
2273
2274         if (vma && vma->vm_flags & VM_NORESERVE)
2275                 return 0;
2276
2277         /*
2278          * Shared mappings base their reservation on the number of pages that
2279          * are already allocated on behalf of the file. Private mappings need
2280          * to reserve the full area even if read-only as mprotect() may be
2281          * called to make the mapping read-write. Assume !vma is a shm mapping
2282          */
2283         if (!vma || vma->vm_flags & VM_SHARED)
2284                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2285         else {
2286                 struct resv_map *resv_map = resv_map_alloc();
2287                 if (!resv_map)
2288                         return -ENOMEM;
2289
2290                 chg = to - from;
2291
2292                 set_vma_resv_map(vma, resv_map);
2293                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2294         }
2295
2296         if (chg < 0)
2297                 return chg;
2298
2299         if (hugetlb_get_quota(inode->i_mapping, chg))
2300                 return -ENOSPC;
2301         ret = hugetlb_acct_memory(h, chg);
2302         if (ret < 0) {
2303                 hugetlb_put_quota(inode->i_mapping, chg);
2304                 return ret;
2305         }
2306         if (!vma || vma->vm_flags & VM_SHARED)
2307                 region_add(&inode->i_mapping->private_list, from, to);
2308         return 0;
2309 }
2310
2311 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2312 {
2313         struct hstate *h = hstate_inode(inode);
2314         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2315
2316         spin_lock(&inode->i_lock);
2317         inode->i_blocks -= blocks_per_huge_page(h);
2318         spin_unlock(&inode->i_lock);
2319
2320         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2321         hugetlb_acct_memory(h, -(chg - freed));
2322 }