mm: hugetlbfs_pagecache_present
[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 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
238
239 /*
240  * Return the page size being used by the MMU to back a VMA. In the majority
241  * of cases, the page size used by the kernel matches the MMU size. On
242  * architectures where it differs, an architecture-specific version of this
243  * function is required.
244  */
245 #ifndef vma_mmu_pagesize
246 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
247 {
248         return vma_kernel_pagesize(vma);
249 }
250 #endif
251
252 /*
253  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
254  * bits of the reservation map pointer, which are always clear due to
255  * alignment.
256  */
257 #define HPAGE_RESV_OWNER    (1UL << 0)
258 #define HPAGE_RESV_UNMAPPED (1UL << 1)
259 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
260
261 /*
262  * These helpers are used to track how many pages are reserved for
263  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
264  * is guaranteed to have their future faults succeed.
265  *
266  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
267  * the reserve counters are updated with the hugetlb_lock held. It is safe
268  * to reset the VMA at fork() time as it is not in use yet and there is no
269  * chance of the global counters getting corrupted as a result of the values.
270  *
271  * The private mapping reservation is represented in a subtly different
272  * manner to a shared mapping.  A shared mapping has a region map associated
273  * with the underlying file, this region map represents the backing file
274  * pages which have ever had a reservation assigned which this persists even
275  * after the page is instantiated.  A private mapping has a region map
276  * associated with the original mmap which is attached to all VMAs which
277  * reference it, this region map represents those offsets which have consumed
278  * reservation ie. where pages have been instantiated.
279  */
280 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
281 {
282         return (unsigned long)vma->vm_private_data;
283 }
284
285 static void set_vma_private_data(struct vm_area_struct *vma,
286                                                         unsigned long value)
287 {
288         vma->vm_private_data = (void *)value;
289 }
290
291 struct resv_map {
292         struct kref refs;
293         struct list_head regions;
294 };
295
296 static struct resv_map *resv_map_alloc(void)
297 {
298         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
299         if (!resv_map)
300                 return NULL;
301
302         kref_init(&resv_map->refs);
303         INIT_LIST_HEAD(&resv_map->regions);
304
305         return resv_map;
306 }
307
308 static void resv_map_release(struct kref *ref)
309 {
310         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
311
312         /* Clear out any active regions before we release the map. */
313         region_truncate(&resv_map->regions, 0);
314         kfree(resv_map);
315 }
316
317 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
318 {
319         VM_BUG_ON(!is_vm_hugetlb_page(vma));
320         if (!(vma->vm_flags & VM_MAYSHARE))
321                 return (struct resv_map *)(get_vma_private_data(vma) &
322                                                         ~HPAGE_RESV_MASK);
323         return NULL;
324 }
325
326 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
327 {
328         VM_BUG_ON(!is_vm_hugetlb_page(vma));
329         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
330
331         set_vma_private_data(vma, (get_vma_private_data(vma) &
332                                 HPAGE_RESV_MASK) | (unsigned long)map);
333 }
334
335 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
336 {
337         VM_BUG_ON(!is_vm_hugetlb_page(vma));
338         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
339
340         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
341 }
342
343 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
344 {
345         VM_BUG_ON(!is_vm_hugetlb_page(vma));
346
347         return (get_vma_private_data(vma) & flag) != 0;
348 }
349
350 /* Decrement the reserved pages in the hugepage pool by one */
351 static void decrement_hugepage_resv_vma(struct hstate *h,
352                         struct vm_area_struct *vma)
353 {
354         if (vma->vm_flags & VM_NORESERVE)
355                 return;
356
357         if (vma->vm_flags & VM_MAYSHARE) {
358                 /* Shared mappings always use reserves */
359                 h->resv_huge_pages--;
360         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
361                 /*
362                  * Only the process that called mmap() has reserves for
363                  * private mappings.
364                  */
365                 h->resv_huge_pages--;
366         }
367 }
368
369 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
370 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
371 {
372         VM_BUG_ON(!is_vm_hugetlb_page(vma));
373         if (!(vma->vm_flags & VM_MAYSHARE))
374                 vma->vm_private_data = (void *)0;
375 }
376
377 /* Returns true if the VMA has associated reserve pages */
378 static int vma_has_reserves(struct vm_area_struct *vma)
379 {
380         if (vma->vm_flags & VM_MAYSHARE)
381                 return 1;
382         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
383                 return 1;
384         return 0;
385 }
386
387 static void clear_gigantic_page(struct page *page,
388                         unsigned long addr, unsigned long sz)
389 {
390         int i;
391         struct page *p = page;
392
393         might_sleep();
394         for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
395                 cond_resched();
396                 clear_user_highpage(p, addr + i * PAGE_SIZE);
397         }
398 }
399 static void clear_huge_page(struct page *page,
400                         unsigned long addr, unsigned long sz)
401 {
402         int i;
403
404         if (unlikely(sz > MAX_ORDER_NR_PAGES)) {
405                 clear_gigantic_page(page, addr, sz);
406                 return;
407         }
408
409         might_sleep();
410         for (i = 0; i < sz/PAGE_SIZE; i++) {
411                 cond_resched();
412                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
413         }
414 }
415
416 static void copy_gigantic_page(struct page *dst, struct page *src,
417                            unsigned long addr, struct vm_area_struct *vma)
418 {
419         int i;
420         struct hstate *h = hstate_vma(vma);
421         struct page *dst_base = dst;
422         struct page *src_base = src;
423         might_sleep();
424         for (i = 0; i < pages_per_huge_page(h); ) {
425                 cond_resched();
426                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
427
428                 i++;
429                 dst = mem_map_next(dst, dst_base, i);
430                 src = mem_map_next(src, src_base, i);
431         }
432 }
433 static void copy_huge_page(struct page *dst, struct page *src,
434                            unsigned long addr, struct vm_area_struct *vma)
435 {
436         int i;
437         struct hstate *h = hstate_vma(vma);
438
439         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
440                 copy_gigantic_page(dst, src, addr, vma);
441                 return;
442         }
443
444         might_sleep();
445         for (i = 0; i < pages_per_huge_page(h); i++) {
446                 cond_resched();
447                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
448         }
449 }
450
451 static void enqueue_huge_page(struct hstate *h, struct page *page)
452 {
453         int nid = page_to_nid(page);
454         list_add(&page->lru, &h->hugepage_freelists[nid]);
455         h->free_huge_pages++;
456         h->free_huge_pages_node[nid]++;
457 }
458
459 static struct page *dequeue_huge_page_vma(struct hstate *h,
460                                 struct vm_area_struct *vma,
461                                 unsigned long address, int avoid_reserve)
462 {
463         int nid;
464         struct page *page = NULL;
465         struct mempolicy *mpol;
466         nodemask_t *nodemask;
467         struct zonelist *zonelist = huge_zonelist(vma, address,
468                                         htlb_alloc_mask, &mpol, &nodemask);
469         struct zone *zone;
470         struct zoneref *z;
471
472         /*
473          * A child process with MAP_PRIVATE mappings created by their parent
474          * have no page reserves. This check ensures that reservations are
475          * not "stolen". The child may still get SIGKILLed
476          */
477         if (!vma_has_reserves(vma) &&
478                         h->free_huge_pages - h->resv_huge_pages == 0)
479                 return NULL;
480
481         /* If reserves cannot be used, ensure enough pages are in the pool */
482         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
483                 return NULL;
484
485         for_each_zone_zonelist_nodemask(zone, z, zonelist,
486                                                 MAX_NR_ZONES - 1, nodemask) {
487                 nid = zone_to_nid(zone);
488                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
489                     !list_empty(&h->hugepage_freelists[nid])) {
490                         page = list_entry(h->hugepage_freelists[nid].next,
491                                           struct page, lru);
492                         list_del(&page->lru);
493                         h->free_huge_pages--;
494                         h->free_huge_pages_node[nid]--;
495
496                         if (!avoid_reserve)
497                                 decrement_hugepage_resv_vma(h, vma);
498
499                         break;
500                 }
501         }
502         mpol_cond_put(mpol);
503         return page;
504 }
505
506 static void update_and_free_page(struct hstate *h, struct page *page)
507 {
508         int i;
509
510         VM_BUG_ON(h->order >= MAX_ORDER);
511
512         h->nr_huge_pages--;
513         h->nr_huge_pages_node[page_to_nid(page)]--;
514         for (i = 0; i < pages_per_huge_page(h); i++) {
515                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
516                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
517                                 1 << PG_private | 1<< PG_writeback);
518         }
519         set_compound_page_dtor(page, NULL);
520         set_page_refcounted(page);
521         arch_release_hugepage(page);
522         __free_pages(page, huge_page_order(h));
523 }
524
525 struct hstate *size_to_hstate(unsigned long size)
526 {
527         struct hstate *h;
528
529         for_each_hstate(h) {
530                 if (huge_page_size(h) == size)
531                         return h;
532         }
533         return NULL;
534 }
535
536 static void free_huge_page(struct page *page)
537 {
538         /*
539          * Can't pass hstate in here because it is called from the
540          * compound page destructor.
541          */
542         struct hstate *h = page_hstate(page);
543         int nid = page_to_nid(page);
544         struct address_space *mapping;
545
546         mapping = (struct address_space *) page_private(page);
547         set_page_private(page, 0);
548         BUG_ON(page_count(page));
549         INIT_LIST_HEAD(&page->lru);
550
551         spin_lock(&hugetlb_lock);
552         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
553                 update_and_free_page(h, page);
554                 h->surplus_huge_pages--;
555                 h->surplus_huge_pages_node[nid]--;
556         } else {
557                 enqueue_huge_page(h, page);
558         }
559         spin_unlock(&hugetlb_lock);
560         if (mapping)
561                 hugetlb_put_quota(mapping, 1);
562 }
563
564 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
565 {
566         set_compound_page_dtor(page, free_huge_page);
567         spin_lock(&hugetlb_lock);
568         h->nr_huge_pages++;
569         h->nr_huge_pages_node[nid]++;
570         spin_unlock(&hugetlb_lock);
571         put_page(page); /* free it into the hugepage allocator */
572 }
573
574 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
575 {
576         int i;
577         int nr_pages = 1 << order;
578         struct page *p = page + 1;
579
580         /* we rely on prep_new_huge_page to set the destructor */
581         set_compound_order(page, order);
582         __SetPageHead(page);
583         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
584                 __SetPageTail(p);
585                 p->first_page = page;
586         }
587 }
588
589 int PageHuge(struct page *page)
590 {
591         compound_page_dtor *dtor;
592
593         if (!PageCompound(page))
594                 return 0;
595
596         page = compound_head(page);
597         dtor = get_compound_page_dtor(page);
598
599         return dtor == free_huge_page;
600 }
601
602 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
603 {
604         struct page *page;
605
606         if (h->order >= MAX_ORDER)
607                 return NULL;
608
609         page = alloc_pages_exact_node(nid,
610                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
611                                                 __GFP_REPEAT|__GFP_NOWARN,
612                 huge_page_order(h));
613         if (page) {
614                 if (arch_prepare_hugepage(page)) {
615                         __free_pages(page, huge_page_order(h));
616                         return NULL;
617                 }
618                 prep_new_huge_page(h, page, nid);
619         }
620
621         return page;
622 }
623
624 /*
625  * Use a helper variable to find the next node and then
626  * copy it back to next_nid_to_alloc afterwards:
627  * otherwise there's a window in which a racer might
628  * pass invalid nid MAX_NUMNODES to alloc_pages_exact_node.
629  * But we don't need to use a spin_lock here: it really
630  * doesn't matter if occasionally a racer chooses the
631  * same nid as we do.  Move nid forward in the mask even
632  * if we just successfully allocated a hugepage so that
633  * the next caller gets hugepages on the next node.
634  */
635 static int hstate_next_node_to_alloc(struct hstate *h)
636 {
637         int next_nid;
638         next_nid = next_node(h->next_nid_to_alloc, node_online_map);
639         if (next_nid == MAX_NUMNODES)
640                 next_nid = first_node(node_online_map);
641         h->next_nid_to_alloc = next_nid;
642         return next_nid;
643 }
644
645 static int alloc_fresh_huge_page(struct hstate *h)
646 {
647         struct page *page;
648         int start_nid;
649         int next_nid;
650         int ret = 0;
651
652         start_nid = h->next_nid_to_alloc;
653         next_nid = start_nid;
654
655         do {
656                 page = alloc_fresh_huge_page_node(h, next_nid);
657                 if (page)
658                         ret = 1;
659                 next_nid = hstate_next_node_to_alloc(h);
660         } while (!page && next_nid != start_nid);
661
662         if (ret)
663                 count_vm_event(HTLB_BUDDY_PGALLOC);
664         else
665                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
666
667         return ret;
668 }
669
670 /*
671  * helper for free_pool_huge_page() - find next node
672  * from which to free a huge page
673  */
674 static int hstate_next_node_to_free(struct hstate *h)
675 {
676         int next_nid;
677         next_nid = next_node(h->next_nid_to_free, node_online_map);
678         if (next_nid == MAX_NUMNODES)
679                 next_nid = first_node(node_online_map);
680         h->next_nid_to_free = next_nid;
681         return next_nid;
682 }
683
684 /*
685  * Free huge page from pool from next node to free.
686  * Attempt to keep persistent huge pages more or less
687  * balanced over allowed nodes.
688  * Called with hugetlb_lock locked.
689  */
690 static int free_pool_huge_page(struct hstate *h, bool acct_surplus)
691 {
692         int start_nid;
693         int next_nid;
694         int ret = 0;
695
696         start_nid = h->next_nid_to_free;
697         next_nid = start_nid;
698
699         do {
700                 /*
701                  * If we're returning unused surplus pages, only examine
702                  * nodes with surplus pages.
703                  */
704                 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
705                     !list_empty(&h->hugepage_freelists[next_nid])) {
706                         struct page *page =
707                                 list_entry(h->hugepage_freelists[next_nid].next,
708                                           struct page, lru);
709                         list_del(&page->lru);
710                         h->free_huge_pages--;
711                         h->free_huge_pages_node[next_nid]--;
712                         if (acct_surplus) {
713                                 h->surplus_huge_pages--;
714                                 h->surplus_huge_pages_node[next_nid]--;
715                         }
716                         update_and_free_page(h, page);
717                         ret = 1;
718                 }
719                 next_nid = hstate_next_node_to_free(h);
720         } while (!ret && next_nid != start_nid);
721
722         return ret;
723 }
724
725 static struct page *alloc_buddy_huge_page(struct hstate *h,
726                         struct vm_area_struct *vma, unsigned long address)
727 {
728         struct page *page;
729         unsigned int nid;
730
731         if (h->order >= MAX_ORDER)
732                 return NULL;
733
734         /*
735          * Assume we will successfully allocate the surplus page to
736          * prevent racing processes from causing the surplus to exceed
737          * overcommit
738          *
739          * This however introduces a different race, where a process B
740          * tries to grow the static hugepage pool while alloc_pages() is
741          * called by process A. B will only examine the per-node
742          * counters in determining if surplus huge pages can be
743          * converted to normal huge pages in adjust_pool_surplus(). A
744          * won't be able to increment the per-node counter, until the
745          * lock is dropped by B, but B doesn't drop hugetlb_lock until
746          * no more huge pages can be converted from surplus to normal
747          * state (and doesn't try to convert again). Thus, we have a
748          * case where a surplus huge page exists, the pool is grown, and
749          * the surplus huge page still exists after, even though it
750          * should just have been converted to a normal huge page. This
751          * does not leak memory, though, as the hugepage will be freed
752          * once it is out of use. It also does not allow the counters to
753          * go out of whack in adjust_pool_surplus() as we don't modify
754          * the node values until we've gotten the hugepage and only the
755          * per-node value is checked there.
756          */
757         spin_lock(&hugetlb_lock);
758         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
759                 spin_unlock(&hugetlb_lock);
760                 return NULL;
761         } else {
762                 h->nr_huge_pages++;
763                 h->surplus_huge_pages++;
764         }
765         spin_unlock(&hugetlb_lock);
766
767         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
768                                         __GFP_REPEAT|__GFP_NOWARN,
769                                         huge_page_order(h));
770
771         if (page && arch_prepare_hugepage(page)) {
772                 __free_pages(page, huge_page_order(h));
773                 return NULL;
774         }
775
776         spin_lock(&hugetlb_lock);
777         if (page) {
778                 /*
779                  * This page is now managed by the hugetlb allocator and has
780                  * no users -- drop the buddy allocator's reference.
781                  */
782                 put_page_testzero(page);
783                 VM_BUG_ON(page_count(page));
784                 nid = page_to_nid(page);
785                 set_compound_page_dtor(page, free_huge_page);
786                 /*
787                  * We incremented the global counters already
788                  */
789                 h->nr_huge_pages_node[nid]++;
790                 h->surplus_huge_pages_node[nid]++;
791                 __count_vm_event(HTLB_BUDDY_PGALLOC);
792         } else {
793                 h->nr_huge_pages--;
794                 h->surplus_huge_pages--;
795                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
796         }
797         spin_unlock(&hugetlb_lock);
798
799         return page;
800 }
801
802 /*
803  * Increase the hugetlb pool such that it can accomodate a reservation
804  * of size 'delta'.
805  */
806 static int gather_surplus_pages(struct hstate *h, int delta)
807 {
808         struct list_head surplus_list;
809         struct page *page, *tmp;
810         int ret, i;
811         int needed, allocated;
812
813         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
814         if (needed <= 0) {
815                 h->resv_huge_pages += delta;
816                 return 0;
817         }
818
819         allocated = 0;
820         INIT_LIST_HEAD(&surplus_list);
821
822         ret = -ENOMEM;
823 retry:
824         spin_unlock(&hugetlb_lock);
825         for (i = 0; i < needed; i++) {
826                 page = alloc_buddy_huge_page(h, NULL, 0);
827                 if (!page) {
828                         /*
829                          * We were not able to allocate enough pages to
830                          * satisfy the entire reservation so we free what
831                          * we've allocated so far.
832                          */
833                         spin_lock(&hugetlb_lock);
834                         needed = 0;
835                         goto free;
836                 }
837
838                 list_add(&page->lru, &surplus_list);
839         }
840         allocated += needed;
841
842         /*
843          * After retaking hugetlb_lock, we need to recalculate 'needed'
844          * because either resv_huge_pages or free_huge_pages may have changed.
845          */
846         spin_lock(&hugetlb_lock);
847         needed = (h->resv_huge_pages + delta) -
848                         (h->free_huge_pages + allocated);
849         if (needed > 0)
850                 goto retry;
851
852         /*
853          * The surplus_list now contains _at_least_ the number of extra pages
854          * needed to accomodate the reservation.  Add the appropriate number
855          * of pages to the hugetlb pool and free the extras back to the buddy
856          * allocator.  Commit the entire reservation here to prevent another
857          * process from stealing the pages as they are added to the pool but
858          * before they are reserved.
859          */
860         needed += allocated;
861         h->resv_huge_pages += delta;
862         ret = 0;
863 free:
864         /* Free the needed pages to the hugetlb pool */
865         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
866                 if ((--needed) < 0)
867                         break;
868                 list_del(&page->lru);
869                 enqueue_huge_page(h, page);
870         }
871
872         /* Free unnecessary surplus pages to the buddy allocator */
873         if (!list_empty(&surplus_list)) {
874                 spin_unlock(&hugetlb_lock);
875                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
876                         list_del(&page->lru);
877                         /*
878                          * The page has a reference count of zero already, so
879                          * call free_huge_page directly instead of using
880                          * put_page.  This must be done with hugetlb_lock
881                          * unlocked which is safe because free_huge_page takes
882                          * hugetlb_lock before deciding how to free the page.
883                          */
884                         free_huge_page(page);
885                 }
886                 spin_lock(&hugetlb_lock);
887         }
888
889         return ret;
890 }
891
892 /*
893  * When releasing a hugetlb pool reservation, any surplus pages that were
894  * allocated to satisfy the reservation must be explicitly freed if they were
895  * never used.
896  * Called with hugetlb_lock held.
897  */
898 static void return_unused_surplus_pages(struct hstate *h,
899                                         unsigned long unused_resv_pages)
900 {
901         unsigned long nr_pages;
902
903         /* Uncommit the reservation */
904         h->resv_huge_pages -= unused_resv_pages;
905
906         /* Cannot return gigantic pages currently */
907         if (h->order >= MAX_ORDER)
908                 return;
909
910         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
911
912         /*
913          * We want to release as many surplus pages as possible, spread
914          * evenly across all nodes. Iterate across all nodes until we
915          * can no longer free unreserved surplus pages. This occurs when
916          * the nodes with surplus pages have no free pages.
917          * free_pool_huge_page() will balance the the frees across the
918          * on-line nodes for us and will handle the hstate accounting.
919          */
920         while (nr_pages--) {
921                 if (!free_pool_huge_page(h, 1))
922                         break;
923         }
924 }
925
926 /*
927  * Determine if the huge page at addr within the vma has an associated
928  * reservation.  Where it does not we will need to logically increase
929  * reservation and actually increase quota before an allocation can occur.
930  * Where any new reservation would be required the reservation change is
931  * prepared, but not committed.  Once the page has been quota'd allocated
932  * an instantiated the change should be committed via vma_commit_reservation.
933  * No action is required on failure.
934  */
935 static long vma_needs_reservation(struct hstate *h,
936                         struct vm_area_struct *vma, unsigned long addr)
937 {
938         struct address_space *mapping = vma->vm_file->f_mapping;
939         struct inode *inode = mapping->host;
940
941         if (vma->vm_flags & VM_MAYSHARE) {
942                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
943                 return region_chg(&inode->i_mapping->private_list,
944                                                         idx, idx + 1);
945
946         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
947                 return 1;
948
949         } else  {
950                 long err;
951                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
952                 struct resv_map *reservations = vma_resv_map(vma);
953
954                 err = region_chg(&reservations->regions, idx, idx + 1);
955                 if (err < 0)
956                         return err;
957                 return 0;
958         }
959 }
960 static void vma_commit_reservation(struct hstate *h,
961                         struct vm_area_struct *vma, unsigned long addr)
962 {
963         struct address_space *mapping = vma->vm_file->f_mapping;
964         struct inode *inode = mapping->host;
965
966         if (vma->vm_flags & VM_MAYSHARE) {
967                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
968                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
969
970         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
971                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
972                 struct resv_map *reservations = vma_resv_map(vma);
973
974                 /* Mark this page used in the map. */
975                 region_add(&reservations->regions, idx, idx + 1);
976         }
977 }
978
979 static struct page *alloc_huge_page(struct vm_area_struct *vma,
980                                     unsigned long addr, int avoid_reserve)
981 {
982         struct hstate *h = hstate_vma(vma);
983         struct page *page;
984         struct address_space *mapping = vma->vm_file->f_mapping;
985         struct inode *inode = mapping->host;
986         long chg;
987
988         /*
989          * Processes that did not create the mapping will have no reserves and
990          * will not have accounted against quota. Check that the quota can be
991          * made before satisfying the allocation
992          * MAP_NORESERVE mappings may also need pages and quota allocated
993          * if no reserve mapping overlaps.
994          */
995         chg = vma_needs_reservation(h, vma, addr);
996         if (chg < 0)
997                 return ERR_PTR(chg);
998         if (chg)
999                 if (hugetlb_get_quota(inode->i_mapping, chg))
1000                         return ERR_PTR(-ENOSPC);
1001
1002         spin_lock(&hugetlb_lock);
1003         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1004         spin_unlock(&hugetlb_lock);
1005
1006         if (!page) {
1007                 page = alloc_buddy_huge_page(h, vma, addr);
1008                 if (!page) {
1009                         hugetlb_put_quota(inode->i_mapping, chg);
1010                         return ERR_PTR(-VM_FAULT_OOM);
1011                 }
1012         }
1013
1014         set_page_refcounted(page);
1015         set_page_private(page, (unsigned long) mapping);
1016
1017         vma_commit_reservation(h, vma, addr);
1018
1019         return page;
1020 }
1021
1022 int __weak alloc_bootmem_huge_page(struct hstate *h)
1023 {
1024         struct huge_bootmem_page *m;
1025         int nr_nodes = nodes_weight(node_online_map);
1026
1027         while (nr_nodes) {
1028                 void *addr;
1029
1030                 addr = __alloc_bootmem_node_nopanic(
1031                                 NODE_DATA(h->next_nid_to_alloc),
1032                                 huge_page_size(h), huge_page_size(h), 0);
1033
1034                 hstate_next_node_to_alloc(h);
1035                 if (addr) {
1036                         /*
1037                          * Use the beginning of the huge page to store the
1038                          * huge_bootmem_page struct (until gather_bootmem
1039                          * puts them into the mem_map).
1040                          */
1041                         m = addr;
1042                         goto found;
1043                 }
1044                 nr_nodes--;
1045         }
1046         return 0;
1047
1048 found:
1049         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1050         /* Put them into a private list first because mem_map is not up yet */
1051         list_add(&m->list, &huge_boot_pages);
1052         m->hstate = h;
1053         return 1;
1054 }
1055
1056 static void prep_compound_huge_page(struct page *page, int order)
1057 {
1058         if (unlikely(order > (MAX_ORDER - 1)))
1059                 prep_compound_gigantic_page(page, order);
1060         else
1061                 prep_compound_page(page, order);
1062 }
1063
1064 /* Put bootmem huge pages into the standard lists after mem_map is up */
1065 static void __init gather_bootmem_prealloc(void)
1066 {
1067         struct huge_bootmem_page *m;
1068
1069         list_for_each_entry(m, &huge_boot_pages, list) {
1070                 struct page *page = virt_to_page(m);
1071                 struct hstate *h = m->hstate;
1072                 __ClearPageReserved(page);
1073                 WARN_ON(page_count(page) != 1);
1074                 prep_compound_huge_page(page, h->order);
1075                 prep_new_huge_page(h, page, page_to_nid(page));
1076         }
1077 }
1078
1079 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1080 {
1081         unsigned long i;
1082
1083         for (i = 0; i < h->max_huge_pages; ++i) {
1084                 if (h->order >= MAX_ORDER) {
1085                         if (!alloc_bootmem_huge_page(h))
1086                                 break;
1087                 } else if (!alloc_fresh_huge_page(h))
1088                         break;
1089         }
1090         h->max_huge_pages = i;
1091 }
1092
1093 static void __init hugetlb_init_hstates(void)
1094 {
1095         struct hstate *h;
1096
1097         for_each_hstate(h) {
1098                 /* oversize hugepages were init'ed in early boot */
1099                 if (h->order < MAX_ORDER)
1100                         hugetlb_hstate_alloc_pages(h);
1101         }
1102 }
1103
1104 static char * __init memfmt(char *buf, unsigned long n)
1105 {
1106         if (n >= (1UL << 30))
1107                 sprintf(buf, "%lu GB", n >> 30);
1108         else if (n >= (1UL << 20))
1109                 sprintf(buf, "%lu MB", n >> 20);
1110         else
1111                 sprintf(buf, "%lu KB", n >> 10);
1112         return buf;
1113 }
1114
1115 static void __init report_hugepages(void)
1116 {
1117         struct hstate *h;
1118
1119         for_each_hstate(h) {
1120                 char buf[32];
1121                 printk(KERN_INFO "HugeTLB registered %s page size, "
1122                                  "pre-allocated %ld pages\n",
1123                         memfmt(buf, huge_page_size(h)),
1124                         h->free_huge_pages);
1125         }
1126 }
1127
1128 #ifdef CONFIG_HIGHMEM
1129 static void try_to_free_low(struct hstate *h, unsigned long count)
1130 {
1131         int i;
1132
1133         if (h->order >= MAX_ORDER)
1134                 return;
1135
1136         for (i = 0; i < MAX_NUMNODES; ++i) {
1137                 struct page *page, *next;
1138                 struct list_head *freel = &h->hugepage_freelists[i];
1139                 list_for_each_entry_safe(page, next, freel, lru) {
1140                         if (count >= h->nr_huge_pages)
1141                                 return;
1142                         if (PageHighMem(page))
1143                                 continue;
1144                         list_del(&page->lru);
1145                         update_and_free_page(h, page);
1146                         h->free_huge_pages--;
1147                         h->free_huge_pages_node[page_to_nid(page)]--;
1148                 }
1149         }
1150 }
1151 #else
1152 static inline void try_to_free_low(struct hstate *h, unsigned long count)
1153 {
1154 }
1155 #endif
1156
1157 /*
1158  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1159  * balanced by operating on them in a round-robin fashion.
1160  * Returns 1 if an adjustment was made.
1161  */
1162 static int adjust_pool_surplus(struct hstate *h, int delta)
1163 {
1164         int start_nid, next_nid;
1165         int ret = 0;
1166
1167         VM_BUG_ON(delta != -1 && delta != 1);
1168
1169         if (delta < 0)
1170                 start_nid = h->next_nid_to_alloc;
1171         else
1172                 start_nid = h->next_nid_to_free;
1173         next_nid = start_nid;
1174
1175         do {
1176                 int nid = next_nid;
1177                 if (delta < 0)  {
1178                         next_nid = hstate_next_node_to_alloc(h);
1179                         /*
1180                          * To shrink on this node, there must be a surplus page
1181                          */
1182                         if (!h->surplus_huge_pages_node[nid])
1183                                 continue;
1184                 }
1185                 if (delta > 0) {
1186                         next_nid = hstate_next_node_to_free(h);
1187                         /*
1188                          * Surplus cannot exceed the total number of pages
1189                          */
1190                         if (h->surplus_huge_pages_node[nid] >=
1191                                                 h->nr_huge_pages_node[nid])
1192                                 continue;
1193                 }
1194
1195                 h->surplus_huge_pages += delta;
1196                 h->surplus_huge_pages_node[nid] += delta;
1197                 ret = 1;
1198                 break;
1199         } while (next_nid != start_nid);
1200
1201         return ret;
1202 }
1203
1204 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1205 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
1206 {
1207         unsigned long min_count, ret;
1208
1209         if (h->order >= MAX_ORDER)
1210                 return h->max_huge_pages;
1211
1212         /*
1213          * Increase the pool size
1214          * First take pages out of surplus state.  Then make up the
1215          * remaining difference by allocating fresh huge pages.
1216          *
1217          * We might race with alloc_buddy_huge_page() here and be unable
1218          * to convert a surplus huge page to a normal huge page. That is
1219          * not critical, though, it just means the overall size of the
1220          * pool might be one hugepage larger than it needs to be, but
1221          * within all the constraints specified by the sysctls.
1222          */
1223         spin_lock(&hugetlb_lock);
1224         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1225                 if (!adjust_pool_surplus(h, -1))
1226                         break;
1227         }
1228
1229         while (count > persistent_huge_pages(h)) {
1230                 /*
1231                  * If this allocation races such that we no longer need the
1232                  * page, free_huge_page will handle it by freeing the page
1233                  * and reducing the surplus.
1234                  */
1235                 spin_unlock(&hugetlb_lock);
1236                 ret = alloc_fresh_huge_page(h);
1237                 spin_lock(&hugetlb_lock);
1238                 if (!ret)
1239                         goto out;
1240
1241         }
1242
1243         /*
1244          * Decrease the pool size
1245          * First return free pages to the buddy allocator (being careful
1246          * to keep enough around to satisfy reservations).  Then place
1247          * pages into surplus state as needed so the pool will shrink
1248          * to the desired size as pages become free.
1249          *
1250          * By placing pages into the surplus state independent of the
1251          * overcommit value, we are allowing the surplus pool size to
1252          * exceed overcommit. There are few sane options here. Since
1253          * alloc_buddy_huge_page() is checking the global counter,
1254          * though, we'll note that we're not allowed to exceed surplus
1255          * and won't grow the pool anywhere else. Not until one of the
1256          * sysctls are changed, or the surplus pages go out of use.
1257          */
1258         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1259         min_count = max(count, min_count);
1260         try_to_free_low(h, min_count);
1261         while (min_count < persistent_huge_pages(h)) {
1262                 if (!free_pool_huge_page(h, 0))
1263                         break;
1264         }
1265         while (count < persistent_huge_pages(h)) {
1266                 if (!adjust_pool_surplus(h, 1))
1267                         break;
1268         }
1269 out:
1270         ret = persistent_huge_pages(h);
1271         spin_unlock(&hugetlb_lock);
1272         return ret;
1273 }
1274
1275 #define HSTATE_ATTR_RO(_name) \
1276         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1277
1278 #define HSTATE_ATTR(_name) \
1279         static struct kobj_attribute _name##_attr = \
1280                 __ATTR(_name, 0644, _name##_show, _name##_store)
1281
1282 static struct kobject *hugepages_kobj;
1283 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1284
1285 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1286 {
1287         int i;
1288         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1289                 if (hstate_kobjs[i] == kobj)
1290                         return &hstates[i];
1291         BUG();
1292         return NULL;
1293 }
1294
1295 static ssize_t nr_hugepages_show(struct kobject *kobj,
1296                                         struct kobj_attribute *attr, char *buf)
1297 {
1298         struct hstate *h = kobj_to_hstate(kobj);
1299         return sprintf(buf, "%lu\n", h->nr_huge_pages);
1300 }
1301 static ssize_t nr_hugepages_store(struct kobject *kobj,
1302                 struct kobj_attribute *attr, const char *buf, size_t count)
1303 {
1304         int err;
1305         unsigned long input;
1306         struct hstate *h = kobj_to_hstate(kobj);
1307
1308         err = strict_strtoul(buf, 10, &input);
1309         if (err)
1310                 return 0;
1311
1312         h->max_huge_pages = set_max_huge_pages(h, input);
1313
1314         return count;
1315 }
1316 HSTATE_ATTR(nr_hugepages);
1317
1318 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1319                                         struct kobj_attribute *attr, char *buf)
1320 {
1321         struct hstate *h = kobj_to_hstate(kobj);
1322         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1323 }
1324 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1325                 struct kobj_attribute *attr, const char *buf, size_t count)
1326 {
1327         int err;
1328         unsigned long input;
1329         struct hstate *h = kobj_to_hstate(kobj);
1330
1331         err = strict_strtoul(buf, 10, &input);
1332         if (err)
1333                 return 0;
1334
1335         spin_lock(&hugetlb_lock);
1336         h->nr_overcommit_huge_pages = input;
1337         spin_unlock(&hugetlb_lock);
1338
1339         return count;
1340 }
1341 HSTATE_ATTR(nr_overcommit_hugepages);
1342
1343 static ssize_t free_hugepages_show(struct kobject *kobj,
1344                                         struct kobj_attribute *attr, char *buf)
1345 {
1346         struct hstate *h = kobj_to_hstate(kobj);
1347         return sprintf(buf, "%lu\n", h->free_huge_pages);
1348 }
1349 HSTATE_ATTR_RO(free_hugepages);
1350
1351 static ssize_t resv_hugepages_show(struct kobject *kobj,
1352                                         struct kobj_attribute *attr, char *buf)
1353 {
1354         struct hstate *h = kobj_to_hstate(kobj);
1355         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1356 }
1357 HSTATE_ATTR_RO(resv_hugepages);
1358
1359 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1360                                         struct kobj_attribute *attr, char *buf)
1361 {
1362         struct hstate *h = kobj_to_hstate(kobj);
1363         return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1364 }
1365 HSTATE_ATTR_RO(surplus_hugepages);
1366
1367 static struct attribute *hstate_attrs[] = {
1368         &nr_hugepages_attr.attr,
1369         &nr_overcommit_hugepages_attr.attr,
1370         &free_hugepages_attr.attr,
1371         &resv_hugepages_attr.attr,
1372         &surplus_hugepages_attr.attr,
1373         NULL,
1374 };
1375
1376 static struct attribute_group hstate_attr_group = {
1377         .attrs = hstate_attrs,
1378 };
1379
1380 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1381 {
1382         int retval;
1383
1384         hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1385                                                         hugepages_kobj);
1386         if (!hstate_kobjs[h - hstates])
1387                 return -ENOMEM;
1388
1389         retval = sysfs_create_group(hstate_kobjs[h - hstates],
1390                                                         &hstate_attr_group);
1391         if (retval)
1392                 kobject_put(hstate_kobjs[h - hstates]);
1393
1394         return retval;
1395 }
1396
1397 static void __init hugetlb_sysfs_init(void)
1398 {
1399         struct hstate *h;
1400         int err;
1401
1402         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1403         if (!hugepages_kobj)
1404                 return;
1405
1406         for_each_hstate(h) {
1407                 err = hugetlb_sysfs_add_hstate(h);
1408                 if (err)
1409                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1410                                                                 h->name);
1411         }
1412 }
1413
1414 static void __exit hugetlb_exit(void)
1415 {
1416         struct hstate *h;
1417
1418         for_each_hstate(h) {
1419                 kobject_put(hstate_kobjs[h - hstates]);
1420         }
1421
1422         kobject_put(hugepages_kobj);
1423 }
1424 module_exit(hugetlb_exit);
1425
1426 static int __init hugetlb_init(void)
1427 {
1428         /* Some platform decide whether they support huge pages at boot
1429          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1430          * there is no such support
1431          */
1432         if (HPAGE_SHIFT == 0)
1433                 return 0;
1434
1435         if (!size_to_hstate(default_hstate_size)) {
1436                 default_hstate_size = HPAGE_SIZE;
1437                 if (!size_to_hstate(default_hstate_size))
1438                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1439         }
1440         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1441         if (default_hstate_max_huge_pages)
1442                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1443
1444         hugetlb_init_hstates();
1445
1446         gather_bootmem_prealloc();
1447
1448         report_hugepages();
1449
1450         hugetlb_sysfs_init();
1451
1452         return 0;
1453 }
1454 module_init(hugetlb_init);
1455
1456 /* Should be called on processing a hugepagesz=... option */
1457 void __init hugetlb_add_hstate(unsigned order)
1458 {
1459         struct hstate *h;
1460         unsigned long i;
1461
1462         if (size_to_hstate(PAGE_SIZE << order)) {
1463                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1464                 return;
1465         }
1466         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1467         BUG_ON(order == 0);
1468         h = &hstates[max_hstate++];
1469         h->order = order;
1470         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1471         h->nr_huge_pages = 0;
1472         h->free_huge_pages = 0;
1473         for (i = 0; i < MAX_NUMNODES; ++i)
1474                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1475         h->next_nid_to_alloc = first_node(node_online_map);
1476         h->next_nid_to_free = first_node(node_online_map);
1477         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1478                                         huge_page_size(h)/1024);
1479
1480         parsed_hstate = h;
1481 }
1482
1483 static int __init hugetlb_nrpages_setup(char *s)
1484 {
1485         unsigned long *mhp;
1486         static unsigned long *last_mhp;
1487
1488         /*
1489          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1490          * so this hugepages= parameter goes to the "default hstate".
1491          */
1492         if (!max_hstate)
1493                 mhp = &default_hstate_max_huge_pages;
1494         else
1495                 mhp = &parsed_hstate->max_huge_pages;
1496
1497         if (mhp == last_mhp) {
1498                 printk(KERN_WARNING "hugepages= specified twice without "
1499                         "interleaving hugepagesz=, ignoring\n");
1500                 return 1;
1501         }
1502
1503         if (sscanf(s, "%lu", mhp) <= 0)
1504                 *mhp = 0;
1505
1506         /*
1507          * Global state is always initialized later in hugetlb_init.
1508          * But we need to allocate >= MAX_ORDER hstates here early to still
1509          * use the bootmem allocator.
1510          */
1511         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1512                 hugetlb_hstate_alloc_pages(parsed_hstate);
1513
1514         last_mhp = mhp;
1515
1516         return 1;
1517 }
1518 __setup("hugepages=", hugetlb_nrpages_setup);
1519
1520 static int __init hugetlb_default_setup(char *s)
1521 {
1522         default_hstate_size = memparse(s, &s);
1523         return 1;
1524 }
1525 __setup("default_hugepagesz=", hugetlb_default_setup);
1526
1527 static unsigned int cpuset_mems_nr(unsigned int *array)
1528 {
1529         int node;
1530         unsigned int nr = 0;
1531
1532         for_each_node_mask(node, cpuset_current_mems_allowed)
1533                 nr += array[node];
1534
1535         return nr;
1536 }
1537
1538 #ifdef CONFIG_SYSCTL
1539 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1540                            struct file *file, void __user *buffer,
1541                            size_t *length, loff_t *ppos)
1542 {
1543         struct hstate *h = &default_hstate;
1544         unsigned long tmp;
1545
1546         if (!write)
1547                 tmp = h->max_huge_pages;
1548
1549         table->data = &tmp;
1550         table->maxlen = sizeof(unsigned long);
1551         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1552
1553         if (write)
1554                 h->max_huge_pages = set_max_huge_pages(h, tmp);
1555
1556         return 0;
1557 }
1558
1559 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1560                         struct file *file, void __user *buffer,
1561                         size_t *length, loff_t *ppos)
1562 {
1563         proc_dointvec(table, write, file, buffer, length, ppos);
1564         if (hugepages_treat_as_movable)
1565                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1566         else
1567                 htlb_alloc_mask = GFP_HIGHUSER;
1568         return 0;
1569 }
1570
1571 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1572                         struct file *file, void __user *buffer,
1573                         size_t *length, loff_t *ppos)
1574 {
1575         struct hstate *h = &default_hstate;
1576         unsigned long tmp;
1577
1578         if (!write)
1579                 tmp = h->nr_overcommit_huge_pages;
1580
1581         table->data = &tmp;
1582         table->maxlen = sizeof(unsigned long);
1583         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1584
1585         if (write) {
1586                 spin_lock(&hugetlb_lock);
1587                 h->nr_overcommit_huge_pages = tmp;
1588                 spin_unlock(&hugetlb_lock);
1589         }
1590
1591         return 0;
1592 }
1593
1594 #endif /* CONFIG_SYSCTL */
1595
1596 void hugetlb_report_meminfo(struct seq_file *m)
1597 {
1598         struct hstate *h = &default_hstate;
1599         seq_printf(m,
1600                         "HugePages_Total:   %5lu\n"
1601                         "HugePages_Free:    %5lu\n"
1602                         "HugePages_Rsvd:    %5lu\n"
1603                         "HugePages_Surp:    %5lu\n"
1604                         "Hugepagesize:   %8lu kB\n",
1605                         h->nr_huge_pages,
1606                         h->free_huge_pages,
1607                         h->resv_huge_pages,
1608                         h->surplus_huge_pages,
1609                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1610 }
1611
1612 int hugetlb_report_node_meminfo(int nid, char *buf)
1613 {
1614         struct hstate *h = &default_hstate;
1615         return sprintf(buf,
1616                 "Node %d HugePages_Total: %5u\n"
1617                 "Node %d HugePages_Free:  %5u\n"
1618                 "Node %d HugePages_Surp:  %5u\n",
1619                 nid, h->nr_huge_pages_node[nid],
1620                 nid, h->free_huge_pages_node[nid],
1621                 nid, h->surplus_huge_pages_node[nid]);
1622 }
1623
1624 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1625 unsigned long hugetlb_total_pages(void)
1626 {
1627         struct hstate *h = &default_hstate;
1628         return h->nr_huge_pages * pages_per_huge_page(h);
1629 }
1630
1631 static int hugetlb_acct_memory(struct hstate *h, long delta)
1632 {
1633         int ret = -ENOMEM;
1634
1635         spin_lock(&hugetlb_lock);
1636         /*
1637          * When cpuset is configured, it breaks the strict hugetlb page
1638          * reservation as the accounting is done on a global variable. Such
1639          * reservation is completely rubbish in the presence of cpuset because
1640          * the reservation is not checked against page availability for the
1641          * current cpuset. Application can still potentially OOM'ed by kernel
1642          * with lack of free htlb page in cpuset that the task is in.
1643          * Attempt to enforce strict accounting with cpuset is almost
1644          * impossible (or too ugly) because cpuset is too fluid that
1645          * task or memory node can be dynamically moved between cpusets.
1646          *
1647          * The change of semantics for shared hugetlb mapping with cpuset is
1648          * undesirable. However, in order to preserve some of the semantics,
1649          * we fall back to check against current free page availability as
1650          * a best attempt and hopefully to minimize the impact of changing
1651          * semantics that cpuset has.
1652          */
1653         if (delta > 0) {
1654                 if (gather_surplus_pages(h, delta) < 0)
1655                         goto out;
1656
1657                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1658                         return_unused_surplus_pages(h, delta);
1659                         goto out;
1660                 }
1661         }
1662
1663         ret = 0;
1664         if (delta < 0)
1665                 return_unused_surplus_pages(h, (unsigned long) -delta);
1666
1667 out:
1668         spin_unlock(&hugetlb_lock);
1669         return ret;
1670 }
1671
1672 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1673 {
1674         struct resv_map *reservations = vma_resv_map(vma);
1675
1676         /*
1677          * This new VMA should share its siblings reservation map if present.
1678          * The VMA will only ever have a valid reservation map pointer where
1679          * it is being copied for another still existing VMA.  As that VMA
1680          * has a reference to the reservation map it cannot dissappear until
1681          * after this open call completes.  It is therefore safe to take a
1682          * new reference here without additional locking.
1683          */
1684         if (reservations)
1685                 kref_get(&reservations->refs);
1686 }
1687
1688 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1689 {
1690         struct hstate *h = hstate_vma(vma);
1691         struct resv_map *reservations = vma_resv_map(vma);
1692         unsigned long reserve;
1693         unsigned long start;
1694         unsigned long end;
1695
1696         if (reservations) {
1697                 start = vma_hugecache_offset(h, vma, vma->vm_start);
1698                 end = vma_hugecache_offset(h, vma, vma->vm_end);
1699
1700                 reserve = (end - start) -
1701                         region_count(&reservations->regions, start, end);
1702
1703                 kref_put(&reservations->refs, resv_map_release);
1704
1705                 if (reserve) {
1706                         hugetlb_acct_memory(h, -reserve);
1707                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1708                 }
1709         }
1710 }
1711
1712 /*
1713  * We cannot handle pagefaults against hugetlb pages at all.  They cause
1714  * handle_mm_fault() to try to instantiate regular-sized pages in the
1715  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
1716  * this far.
1717  */
1718 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1719 {
1720         BUG();
1721         return 0;
1722 }
1723
1724 struct vm_operations_struct hugetlb_vm_ops = {
1725         .fault = hugetlb_vm_op_fault,
1726         .open = hugetlb_vm_op_open,
1727         .close = hugetlb_vm_op_close,
1728 };
1729
1730 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1731                                 int writable)
1732 {
1733         pte_t entry;
1734
1735         if (writable) {
1736                 entry =
1737                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1738         } else {
1739                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1740         }
1741         entry = pte_mkyoung(entry);
1742         entry = pte_mkhuge(entry);
1743
1744         return entry;
1745 }
1746
1747 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1748                                    unsigned long address, pte_t *ptep)
1749 {
1750         pte_t entry;
1751
1752         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1753         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1754                 update_mmu_cache(vma, address, entry);
1755         }
1756 }
1757
1758
1759 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1760                             struct vm_area_struct *vma)
1761 {
1762         pte_t *src_pte, *dst_pte, entry;
1763         struct page *ptepage;
1764         unsigned long addr;
1765         int cow;
1766         struct hstate *h = hstate_vma(vma);
1767         unsigned long sz = huge_page_size(h);
1768
1769         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1770
1771         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1772                 src_pte = huge_pte_offset(src, addr);
1773                 if (!src_pte)
1774                         continue;
1775                 dst_pte = huge_pte_alloc(dst, addr, sz);
1776                 if (!dst_pte)
1777                         goto nomem;
1778
1779                 /* If the pagetables are shared don't copy or take references */
1780                 if (dst_pte == src_pte)
1781                         continue;
1782
1783                 spin_lock(&dst->page_table_lock);
1784                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1785                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1786                         if (cow)
1787                                 huge_ptep_set_wrprotect(src, addr, src_pte);
1788                         entry = huge_ptep_get(src_pte);
1789                         ptepage = pte_page(entry);
1790                         get_page(ptepage);
1791                         set_huge_pte_at(dst, addr, dst_pte, entry);
1792                 }
1793                 spin_unlock(&src->page_table_lock);
1794                 spin_unlock(&dst->page_table_lock);
1795         }
1796         return 0;
1797
1798 nomem:
1799         return -ENOMEM;
1800 }
1801
1802 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1803                             unsigned long end, struct page *ref_page)
1804 {
1805         struct mm_struct *mm = vma->vm_mm;
1806         unsigned long address;
1807         pte_t *ptep;
1808         pte_t pte;
1809         struct page *page;
1810         struct page *tmp;
1811         struct hstate *h = hstate_vma(vma);
1812         unsigned long sz = huge_page_size(h);
1813
1814         /*
1815          * A page gathering list, protected by per file i_mmap_lock. The
1816          * lock is used to avoid list corruption from multiple unmapping
1817          * of the same page since we are using page->lru.
1818          */
1819         LIST_HEAD(page_list);
1820
1821         WARN_ON(!is_vm_hugetlb_page(vma));
1822         BUG_ON(start & ~huge_page_mask(h));
1823         BUG_ON(end & ~huge_page_mask(h));
1824
1825         mmu_notifier_invalidate_range_start(mm, start, end);
1826         spin_lock(&mm->page_table_lock);
1827         for (address = start; address < end; address += sz) {
1828                 ptep = huge_pte_offset(mm, address);
1829                 if (!ptep)
1830                         continue;
1831
1832                 if (huge_pmd_unshare(mm, &address, ptep))
1833                         continue;
1834
1835                 /*
1836                  * If a reference page is supplied, it is because a specific
1837                  * page is being unmapped, not a range. Ensure the page we
1838                  * are about to unmap is the actual page of interest.
1839                  */
1840                 if (ref_page) {
1841                         pte = huge_ptep_get(ptep);
1842                         if (huge_pte_none(pte))
1843                                 continue;
1844                         page = pte_page(pte);
1845                         if (page != ref_page)
1846                                 continue;
1847
1848                         /*
1849                          * Mark the VMA as having unmapped its page so that
1850                          * future faults in this VMA will fail rather than
1851                          * looking like data was lost
1852                          */
1853                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1854                 }
1855
1856                 pte = huge_ptep_get_and_clear(mm, address, ptep);
1857                 if (huge_pte_none(pte))
1858                         continue;
1859
1860                 page = pte_page(pte);
1861                 if (pte_dirty(pte))
1862                         set_page_dirty(page);
1863                 list_add(&page->lru, &page_list);
1864         }
1865         spin_unlock(&mm->page_table_lock);
1866         flush_tlb_range(vma, start, end);
1867         mmu_notifier_invalidate_range_end(mm, start, end);
1868         list_for_each_entry_safe(page, tmp, &page_list, lru) {
1869                 list_del(&page->lru);
1870                 put_page(page);
1871         }
1872 }
1873
1874 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1875                           unsigned long end, struct page *ref_page)
1876 {
1877         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1878         __unmap_hugepage_range(vma, start, end, ref_page);
1879         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1880 }
1881
1882 /*
1883  * This is called when the original mapper is failing to COW a MAP_PRIVATE
1884  * mappping it owns the reserve page for. The intention is to unmap the page
1885  * from other VMAs and let the children be SIGKILLed if they are faulting the
1886  * same region.
1887  */
1888 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1889                                 struct page *page, unsigned long address)
1890 {
1891         struct hstate *h = hstate_vma(vma);
1892         struct vm_area_struct *iter_vma;
1893         struct address_space *mapping;
1894         struct prio_tree_iter iter;
1895         pgoff_t pgoff;
1896
1897         /*
1898          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1899          * from page cache lookup which is in HPAGE_SIZE units.
1900          */
1901         address = address & huge_page_mask(h);
1902         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1903                 + (vma->vm_pgoff >> PAGE_SHIFT);
1904         mapping = (struct address_space *)page_private(page);
1905
1906         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1907                 /* Do not unmap the current VMA */
1908                 if (iter_vma == vma)
1909                         continue;
1910
1911                 /*
1912                  * Unmap the page from other VMAs without their own reserves.
1913                  * They get marked to be SIGKILLed if they fault in these
1914                  * areas. This is because a future no-page fault on this VMA
1915                  * could insert a zeroed page instead of the data existing
1916                  * from the time of fork. This would look like data corruption
1917                  */
1918                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1919                         unmap_hugepage_range(iter_vma,
1920                                 address, address + huge_page_size(h),
1921                                 page);
1922         }
1923
1924         return 1;
1925 }
1926
1927 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1928                         unsigned long address, pte_t *ptep, pte_t pte,
1929                         struct page *pagecache_page)
1930 {
1931         struct hstate *h = hstate_vma(vma);
1932         struct page *old_page, *new_page;
1933         int avoidcopy;
1934         int outside_reserve = 0;
1935
1936         old_page = pte_page(pte);
1937
1938 retry_avoidcopy:
1939         /* If no-one else is actually using this page, avoid the copy
1940          * and just make the page writable */
1941         avoidcopy = (page_count(old_page) == 1);
1942         if (avoidcopy) {
1943                 set_huge_ptep_writable(vma, address, ptep);
1944                 return 0;
1945         }
1946
1947         /*
1948          * If the process that created a MAP_PRIVATE mapping is about to
1949          * perform a COW due to a shared page count, attempt to satisfy
1950          * the allocation without using the existing reserves. The pagecache
1951          * page is used to determine if the reserve at this address was
1952          * consumed or not. If reserves were used, a partial faulted mapping
1953          * at the time of fork() could consume its reserves on COW instead
1954          * of the full address range.
1955          */
1956         if (!(vma->vm_flags & VM_MAYSHARE) &&
1957                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1958                         old_page != pagecache_page)
1959                 outside_reserve = 1;
1960
1961         page_cache_get(old_page);
1962         new_page = alloc_huge_page(vma, address, outside_reserve);
1963
1964         if (IS_ERR(new_page)) {
1965                 page_cache_release(old_page);
1966
1967                 /*
1968                  * If a process owning a MAP_PRIVATE mapping fails to COW,
1969                  * it is due to references held by a child and an insufficient
1970                  * huge page pool. To guarantee the original mappers
1971                  * reliability, unmap the page from child processes. The child
1972                  * may get SIGKILLed if it later faults.
1973                  */
1974                 if (outside_reserve) {
1975                         BUG_ON(huge_pte_none(pte));
1976                         if (unmap_ref_private(mm, vma, old_page, address)) {
1977                                 BUG_ON(page_count(old_page) != 1);
1978                                 BUG_ON(huge_pte_none(pte));
1979                                 goto retry_avoidcopy;
1980                         }
1981                         WARN_ON_ONCE(1);
1982                 }
1983
1984                 return -PTR_ERR(new_page);
1985         }
1986
1987         spin_unlock(&mm->page_table_lock);
1988         copy_huge_page(new_page, old_page, address, vma);
1989         __SetPageUptodate(new_page);
1990         spin_lock(&mm->page_table_lock);
1991
1992         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1993         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1994                 /* Break COW */
1995                 huge_ptep_clear_flush(vma, address, ptep);
1996                 set_huge_pte_at(mm, address, ptep,
1997                                 make_huge_pte(vma, new_page, 1));
1998                 /* Make the old page be freed below */
1999                 new_page = old_page;
2000         }
2001         page_cache_release(new_page);
2002         page_cache_release(old_page);
2003         return 0;
2004 }
2005
2006 /* Return the pagecache page at a given address within a VMA */
2007 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2008                         struct vm_area_struct *vma, unsigned long address)
2009 {
2010         struct address_space *mapping;
2011         pgoff_t idx;
2012
2013         mapping = vma->vm_file->f_mapping;
2014         idx = vma_hugecache_offset(h, vma, address);
2015
2016         return find_lock_page(mapping, idx);
2017 }
2018
2019 /*
2020  * Return whether there is a pagecache page to back given address within VMA.
2021  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2022  */
2023 static bool hugetlbfs_pagecache_present(struct hstate *h,
2024                         struct vm_area_struct *vma, unsigned long address)
2025 {
2026         struct address_space *mapping;
2027         pgoff_t idx;
2028         struct page *page;
2029
2030         mapping = vma->vm_file->f_mapping;
2031         idx = vma_hugecache_offset(h, vma, address);
2032
2033         page = find_get_page(mapping, idx);
2034         if (page)
2035                 put_page(page);
2036         return page != NULL;
2037 }
2038
2039 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2040                         unsigned long address, pte_t *ptep, unsigned int flags)
2041 {
2042         struct hstate *h = hstate_vma(vma);
2043         int ret = VM_FAULT_SIGBUS;
2044         pgoff_t idx;
2045         unsigned long size;
2046         struct page *page;
2047         struct address_space *mapping;
2048         pte_t new_pte;
2049
2050         /*
2051          * Currently, we are forced to kill the process in the event the
2052          * original mapper has unmapped pages from the child due to a failed
2053          * COW. Warn that such a situation has occured as it may not be obvious
2054          */
2055         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2056                 printk(KERN_WARNING
2057                         "PID %d killed due to inadequate hugepage pool\n",
2058                         current->pid);
2059                 return ret;
2060         }
2061
2062         mapping = vma->vm_file->f_mapping;
2063         idx = vma_hugecache_offset(h, vma, address);
2064
2065         /*
2066          * Use page lock to guard against racing truncation
2067          * before we get page_table_lock.
2068          */
2069 retry:
2070         page = find_lock_page(mapping, idx);
2071         if (!page) {
2072                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2073                 if (idx >= size)
2074                         goto out;
2075                 page = alloc_huge_page(vma, address, 0);
2076                 if (IS_ERR(page)) {
2077                         ret = -PTR_ERR(page);
2078                         goto out;
2079                 }
2080                 clear_huge_page(page, address, huge_page_size(h));
2081                 __SetPageUptodate(page);
2082
2083                 if (vma->vm_flags & VM_MAYSHARE) {
2084                         int err;
2085                         struct inode *inode = mapping->host;
2086
2087                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2088                         if (err) {
2089                                 put_page(page);
2090                                 if (err == -EEXIST)
2091                                         goto retry;
2092                                 goto out;
2093                         }
2094
2095                         spin_lock(&inode->i_lock);
2096                         inode->i_blocks += blocks_per_huge_page(h);
2097                         spin_unlock(&inode->i_lock);
2098                 } else
2099                         lock_page(page);
2100         }
2101
2102         /*
2103          * If we are going to COW a private mapping later, we examine the
2104          * pending reservations for this page now. This will ensure that
2105          * any allocations necessary to record that reservation occur outside
2106          * the spinlock.
2107          */
2108         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2109                 if (vma_needs_reservation(h, vma, address) < 0) {
2110                         ret = VM_FAULT_OOM;
2111                         goto backout_unlocked;
2112                 }
2113
2114         spin_lock(&mm->page_table_lock);
2115         size = i_size_read(mapping->host) >> huge_page_shift(h);
2116         if (idx >= size)
2117                 goto backout;
2118
2119         ret = 0;
2120         if (!huge_pte_none(huge_ptep_get(ptep)))
2121                 goto backout;
2122
2123         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2124                                 && (vma->vm_flags & VM_SHARED)));
2125         set_huge_pte_at(mm, address, ptep, new_pte);
2126
2127         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2128                 /* Optimization, do the COW without a second fault */
2129                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2130         }
2131
2132         spin_unlock(&mm->page_table_lock);
2133         unlock_page(page);
2134 out:
2135         return ret;
2136
2137 backout:
2138         spin_unlock(&mm->page_table_lock);
2139 backout_unlocked:
2140         unlock_page(page);
2141         put_page(page);
2142         goto out;
2143 }
2144
2145 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2146                         unsigned long address, unsigned int flags)
2147 {
2148         pte_t *ptep;
2149         pte_t entry;
2150         int ret;
2151         struct page *pagecache_page = NULL;
2152         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2153         struct hstate *h = hstate_vma(vma);
2154
2155         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2156         if (!ptep)
2157                 return VM_FAULT_OOM;
2158
2159         /*
2160          * Serialize hugepage allocation and instantiation, so that we don't
2161          * get spurious allocation failures if two CPUs race to instantiate
2162          * the same page in the page cache.
2163          */
2164         mutex_lock(&hugetlb_instantiation_mutex);
2165         entry = huge_ptep_get(ptep);
2166         if (huge_pte_none(entry)) {
2167                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2168                 goto out_mutex;
2169         }
2170
2171         ret = 0;
2172
2173         /*
2174          * If we are going to COW the mapping later, we examine the pending
2175          * reservations for this page now. This will ensure that any
2176          * allocations necessary to record that reservation occur outside the
2177          * spinlock. For private mappings, we also lookup the pagecache
2178          * page now as it is used to determine if a reservation has been
2179          * consumed.
2180          */
2181         if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2182                 if (vma_needs_reservation(h, vma, address) < 0) {
2183                         ret = VM_FAULT_OOM;
2184                         goto out_mutex;
2185                 }
2186
2187                 if (!(vma->vm_flags & VM_MAYSHARE))
2188                         pagecache_page = hugetlbfs_pagecache_page(h,
2189                                                                 vma, address);
2190         }
2191
2192         spin_lock(&mm->page_table_lock);
2193         /* Check for a racing update before calling hugetlb_cow */
2194         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2195                 goto out_page_table_lock;
2196
2197
2198         if (flags & FAULT_FLAG_WRITE) {
2199                 if (!pte_write(entry)) {
2200                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2201                                                         pagecache_page);
2202                         goto out_page_table_lock;
2203                 }
2204                 entry = pte_mkdirty(entry);
2205         }
2206         entry = pte_mkyoung(entry);
2207         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2208                                                 flags & FAULT_FLAG_WRITE))
2209                 update_mmu_cache(vma, address, entry);
2210
2211 out_page_table_lock:
2212         spin_unlock(&mm->page_table_lock);
2213
2214         if (pagecache_page) {
2215                 unlock_page(pagecache_page);
2216                 put_page(pagecache_page);
2217         }
2218
2219 out_mutex:
2220         mutex_unlock(&hugetlb_instantiation_mutex);
2221
2222         return ret;
2223 }
2224
2225 /* Can be overriden by architectures */
2226 __attribute__((weak)) struct page *
2227 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2228                pud_t *pud, int write)
2229 {
2230         BUG();
2231         return NULL;
2232 }
2233
2234 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2235                         struct page **pages, struct vm_area_struct **vmas,
2236                         unsigned long *position, int *length, int i,
2237                         unsigned int flags)
2238 {
2239         unsigned long pfn_offset;
2240         unsigned long vaddr = *position;
2241         int remainder = *length;
2242         struct hstate *h = hstate_vma(vma);
2243
2244         spin_lock(&mm->page_table_lock);
2245         while (vaddr < vma->vm_end && remainder) {
2246                 pte_t *pte;
2247                 int absent;
2248                 struct page *page;
2249
2250                 /*
2251                  * Some archs (sparc64, sh*) have multiple pte_ts to
2252                  * each hugepage.  We have to make sure we get the
2253                  * first, for the page indexing below to work.
2254                  */
2255                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2256                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2257
2258                 /*
2259                  * When coredumping, it suits get_dump_page if we just return
2260                  * an error where there's an empty slot with no huge pagecache
2261                  * to back it.  This way, we avoid allocating a hugepage, and
2262                  * the sparse dumpfile avoids allocating disk blocks, but its
2263                  * huge holes still show up with zeroes where they need to be.
2264                  */
2265                 if (absent && (flags & FOLL_DUMP) &&
2266                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2267                         remainder = 0;
2268                         break;
2269                 }
2270
2271                 if (absent ||
2272                     ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2273                         int ret;
2274
2275                         spin_unlock(&mm->page_table_lock);
2276                         ret = hugetlb_fault(mm, vma, vaddr,
2277                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2278                         spin_lock(&mm->page_table_lock);
2279                         if (!(ret & VM_FAULT_ERROR))
2280                                 continue;
2281
2282                         remainder = 0;
2283                         break;
2284                 }
2285
2286                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2287                 page = pte_page(huge_ptep_get(pte));
2288 same_page:
2289                 if (pages) {
2290                         pages[i] = mem_map_offset(page, pfn_offset);
2291                         get_page(pages[i]);
2292                 }
2293
2294                 if (vmas)
2295                         vmas[i] = vma;
2296
2297                 vaddr += PAGE_SIZE;
2298                 ++pfn_offset;
2299                 --remainder;
2300                 ++i;
2301                 if (vaddr < vma->vm_end && remainder &&
2302                                 pfn_offset < pages_per_huge_page(h)) {
2303                         /*
2304                          * We use pfn_offset to avoid touching the pageframes
2305                          * of this compound page.
2306                          */
2307                         goto same_page;
2308                 }
2309         }
2310         spin_unlock(&mm->page_table_lock);
2311         *length = remainder;
2312         *position = vaddr;
2313
2314         return i ? i : -EFAULT;
2315 }
2316
2317 void hugetlb_change_protection(struct vm_area_struct *vma,
2318                 unsigned long address, unsigned long end, pgprot_t newprot)
2319 {
2320         struct mm_struct *mm = vma->vm_mm;
2321         unsigned long start = address;
2322         pte_t *ptep;
2323         pte_t pte;
2324         struct hstate *h = hstate_vma(vma);
2325
2326         BUG_ON(address >= end);
2327         flush_cache_range(vma, address, end);
2328
2329         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2330         spin_lock(&mm->page_table_lock);
2331         for (; address < end; address += huge_page_size(h)) {
2332                 ptep = huge_pte_offset(mm, address);
2333                 if (!ptep)
2334                         continue;
2335                 if (huge_pmd_unshare(mm, &address, ptep))
2336                         continue;
2337                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2338                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2339                         pte = pte_mkhuge(pte_modify(pte, newprot));
2340                         set_huge_pte_at(mm, address, ptep, pte);
2341                 }
2342         }
2343         spin_unlock(&mm->page_table_lock);
2344         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2345
2346         flush_tlb_range(vma, start, end);
2347 }
2348
2349 int hugetlb_reserve_pages(struct inode *inode,
2350                                         long from, long to,
2351                                         struct vm_area_struct *vma,
2352                                         int acctflag)
2353 {
2354         long ret, chg;
2355         struct hstate *h = hstate_inode(inode);
2356
2357         /*
2358          * Only apply hugepage reservation if asked. At fault time, an
2359          * attempt will be made for VM_NORESERVE to allocate a page
2360          * and filesystem quota without using reserves
2361          */
2362         if (acctflag & VM_NORESERVE)
2363                 return 0;
2364
2365         /*
2366          * Shared mappings base their reservation on the number of pages that
2367          * are already allocated on behalf of the file. Private mappings need
2368          * to reserve the full area even if read-only as mprotect() may be
2369          * called to make the mapping read-write. Assume !vma is a shm mapping
2370          */
2371         if (!vma || vma->vm_flags & VM_MAYSHARE)
2372                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2373         else {
2374                 struct resv_map *resv_map = resv_map_alloc();
2375                 if (!resv_map)
2376                         return -ENOMEM;
2377
2378                 chg = to - from;
2379
2380                 set_vma_resv_map(vma, resv_map);
2381                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2382         }
2383
2384         if (chg < 0)
2385                 return chg;
2386
2387         /* There must be enough filesystem quota for the mapping */
2388         if (hugetlb_get_quota(inode->i_mapping, chg))
2389                 return -ENOSPC;
2390
2391         /*
2392          * Check enough hugepages are available for the reservation.
2393          * Hand back the quota if there are not
2394          */
2395         ret = hugetlb_acct_memory(h, chg);
2396         if (ret < 0) {
2397                 hugetlb_put_quota(inode->i_mapping, chg);
2398                 return ret;
2399         }
2400
2401         /*
2402          * Account for the reservations made. Shared mappings record regions
2403          * that have reservations as they are shared by multiple VMAs.
2404          * When the last VMA disappears, the region map says how much
2405          * the reservation was and the page cache tells how much of
2406          * the reservation was consumed. Private mappings are per-VMA and
2407          * only the consumed reservations are tracked. When the VMA
2408          * disappears, the original reservation is the VMA size and the
2409          * consumed reservations are stored in the map. Hence, nothing
2410          * else has to be done for private mappings here
2411          */
2412         if (!vma || vma->vm_flags & VM_MAYSHARE)
2413                 region_add(&inode->i_mapping->private_list, from, to);
2414         return 0;
2415 }
2416
2417 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2418 {
2419         struct hstate *h = hstate_inode(inode);
2420         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2421
2422         spin_lock(&inode->i_lock);
2423         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2424         spin_unlock(&inode->i_lock);
2425
2426         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2427         hugetlb_acct_memory(h, -(chg - freed));
2428 }