hugetlb: derive huge pages nodes allowed from task mempolicy
[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  * common helper functions for hstate_next_node_to_{alloc|free}.
626  * We may have allocated or freed a huge page based on a different
627  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
628  * be outside of *nodes_allowed.  Ensure that we use an allowed
629  * node for alloc or free.
630  */
631 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
632 {
633         nid = next_node(nid, *nodes_allowed);
634         if (nid == MAX_NUMNODES)
635                 nid = first_node(*nodes_allowed);
636         VM_BUG_ON(nid >= MAX_NUMNODES);
637
638         return nid;
639 }
640
641 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
642 {
643         if (!node_isset(nid, *nodes_allowed))
644                 nid = next_node_allowed(nid, nodes_allowed);
645         return nid;
646 }
647
648 /*
649  * returns the previously saved node ["this node"] from which to
650  * allocate a persistent huge page for the pool and advance the
651  * next node from which to allocate, handling wrap at end of node
652  * mask.
653  */
654 static int hstate_next_node_to_alloc(struct hstate *h,
655                                         nodemask_t *nodes_allowed)
656 {
657         int nid;
658
659         VM_BUG_ON(!nodes_allowed);
660
661         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
662         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
663
664         return nid;
665 }
666
667 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
668 {
669         struct page *page;
670         int start_nid;
671         int next_nid;
672         int ret = 0;
673
674         start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
675         next_nid = start_nid;
676
677         do {
678                 page = alloc_fresh_huge_page_node(h, next_nid);
679                 if (page) {
680                         ret = 1;
681                         break;
682                 }
683                 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
684         } while (next_nid != start_nid);
685
686         if (ret)
687                 count_vm_event(HTLB_BUDDY_PGALLOC);
688         else
689                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
690
691         return ret;
692 }
693
694 /*
695  * helper for free_pool_huge_page() - return the previously saved
696  * node ["this node"] from which to free a huge page.  Advance the
697  * next node id whether or not we find a free huge page to free so
698  * that the next attempt to free addresses the next node.
699  */
700 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
701 {
702         int nid;
703
704         VM_BUG_ON(!nodes_allowed);
705
706         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
707         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
708
709         return nid;
710 }
711
712 /*
713  * Free huge page from pool from next node to free.
714  * Attempt to keep persistent huge pages more or less
715  * balanced over allowed nodes.
716  * Called with hugetlb_lock locked.
717  */
718 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
719                                                          bool acct_surplus)
720 {
721         int start_nid;
722         int next_nid;
723         int ret = 0;
724
725         start_nid = hstate_next_node_to_free(h, nodes_allowed);
726         next_nid = start_nid;
727
728         do {
729                 /*
730                  * If we're returning unused surplus pages, only examine
731                  * nodes with surplus pages.
732                  */
733                 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
734                     !list_empty(&h->hugepage_freelists[next_nid])) {
735                         struct page *page =
736                                 list_entry(h->hugepage_freelists[next_nid].next,
737                                           struct page, lru);
738                         list_del(&page->lru);
739                         h->free_huge_pages--;
740                         h->free_huge_pages_node[next_nid]--;
741                         if (acct_surplus) {
742                                 h->surplus_huge_pages--;
743                                 h->surplus_huge_pages_node[next_nid]--;
744                         }
745                         update_and_free_page(h, page);
746                         ret = 1;
747                         break;
748                 }
749                 next_nid = hstate_next_node_to_free(h, nodes_allowed);
750         } while (next_nid != start_nid);
751
752         return ret;
753 }
754
755 static struct page *alloc_buddy_huge_page(struct hstate *h,
756                         struct vm_area_struct *vma, unsigned long address)
757 {
758         struct page *page;
759         unsigned int nid;
760
761         if (h->order >= MAX_ORDER)
762                 return NULL;
763
764         /*
765          * Assume we will successfully allocate the surplus page to
766          * prevent racing processes from causing the surplus to exceed
767          * overcommit
768          *
769          * This however introduces a different race, where a process B
770          * tries to grow the static hugepage pool while alloc_pages() is
771          * called by process A. B will only examine the per-node
772          * counters in determining if surplus huge pages can be
773          * converted to normal huge pages in adjust_pool_surplus(). A
774          * won't be able to increment the per-node counter, until the
775          * lock is dropped by B, but B doesn't drop hugetlb_lock until
776          * no more huge pages can be converted from surplus to normal
777          * state (and doesn't try to convert again). Thus, we have a
778          * case where a surplus huge page exists, the pool is grown, and
779          * the surplus huge page still exists after, even though it
780          * should just have been converted to a normal huge page. This
781          * does not leak memory, though, as the hugepage will be freed
782          * once it is out of use. It also does not allow the counters to
783          * go out of whack in adjust_pool_surplus() as we don't modify
784          * the node values until we've gotten the hugepage and only the
785          * per-node value is checked there.
786          */
787         spin_lock(&hugetlb_lock);
788         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
789                 spin_unlock(&hugetlb_lock);
790                 return NULL;
791         } else {
792                 h->nr_huge_pages++;
793                 h->surplus_huge_pages++;
794         }
795         spin_unlock(&hugetlb_lock);
796
797         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
798                                         __GFP_REPEAT|__GFP_NOWARN,
799                                         huge_page_order(h));
800
801         if (page && arch_prepare_hugepage(page)) {
802                 __free_pages(page, huge_page_order(h));
803                 return NULL;
804         }
805
806         spin_lock(&hugetlb_lock);
807         if (page) {
808                 /*
809                  * This page is now managed by the hugetlb allocator and has
810                  * no users -- drop the buddy allocator's reference.
811                  */
812                 put_page_testzero(page);
813                 VM_BUG_ON(page_count(page));
814                 nid = page_to_nid(page);
815                 set_compound_page_dtor(page, free_huge_page);
816                 /*
817                  * We incremented the global counters already
818                  */
819                 h->nr_huge_pages_node[nid]++;
820                 h->surplus_huge_pages_node[nid]++;
821                 __count_vm_event(HTLB_BUDDY_PGALLOC);
822         } else {
823                 h->nr_huge_pages--;
824                 h->surplus_huge_pages--;
825                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
826         }
827         spin_unlock(&hugetlb_lock);
828
829         return page;
830 }
831
832 /*
833  * Increase the hugetlb pool such that it can accomodate a reservation
834  * of size 'delta'.
835  */
836 static int gather_surplus_pages(struct hstate *h, int delta)
837 {
838         struct list_head surplus_list;
839         struct page *page, *tmp;
840         int ret, i;
841         int needed, allocated;
842
843         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
844         if (needed <= 0) {
845                 h->resv_huge_pages += delta;
846                 return 0;
847         }
848
849         allocated = 0;
850         INIT_LIST_HEAD(&surplus_list);
851
852         ret = -ENOMEM;
853 retry:
854         spin_unlock(&hugetlb_lock);
855         for (i = 0; i < needed; i++) {
856                 page = alloc_buddy_huge_page(h, NULL, 0);
857                 if (!page) {
858                         /*
859                          * We were not able to allocate enough pages to
860                          * satisfy the entire reservation so we free what
861                          * we've allocated so far.
862                          */
863                         spin_lock(&hugetlb_lock);
864                         needed = 0;
865                         goto free;
866                 }
867
868                 list_add(&page->lru, &surplus_list);
869         }
870         allocated += needed;
871
872         /*
873          * After retaking hugetlb_lock, we need to recalculate 'needed'
874          * because either resv_huge_pages or free_huge_pages may have changed.
875          */
876         spin_lock(&hugetlb_lock);
877         needed = (h->resv_huge_pages + delta) -
878                         (h->free_huge_pages + allocated);
879         if (needed > 0)
880                 goto retry;
881
882         /*
883          * The surplus_list now contains _at_least_ the number of extra pages
884          * needed to accomodate the reservation.  Add the appropriate number
885          * of pages to the hugetlb pool and free the extras back to the buddy
886          * allocator.  Commit the entire reservation here to prevent another
887          * process from stealing the pages as they are added to the pool but
888          * before they are reserved.
889          */
890         needed += allocated;
891         h->resv_huge_pages += delta;
892         ret = 0;
893 free:
894         /* Free the needed pages to the hugetlb pool */
895         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
896                 if ((--needed) < 0)
897                         break;
898                 list_del(&page->lru);
899                 enqueue_huge_page(h, page);
900         }
901
902         /* Free unnecessary surplus pages to the buddy allocator */
903         if (!list_empty(&surplus_list)) {
904                 spin_unlock(&hugetlb_lock);
905                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
906                         list_del(&page->lru);
907                         /*
908                          * The page has a reference count of zero already, so
909                          * call free_huge_page directly instead of using
910                          * put_page.  This must be done with hugetlb_lock
911                          * unlocked which is safe because free_huge_page takes
912                          * hugetlb_lock before deciding how to free the page.
913                          */
914                         free_huge_page(page);
915                 }
916                 spin_lock(&hugetlb_lock);
917         }
918
919         return ret;
920 }
921
922 /*
923  * When releasing a hugetlb pool reservation, any surplus pages that were
924  * allocated to satisfy the reservation must be explicitly freed if they were
925  * never used.
926  * Called with hugetlb_lock held.
927  */
928 static void return_unused_surplus_pages(struct hstate *h,
929                                         unsigned long unused_resv_pages)
930 {
931         unsigned long nr_pages;
932
933         /* Uncommit the reservation */
934         h->resv_huge_pages -= unused_resv_pages;
935
936         /* Cannot return gigantic pages currently */
937         if (h->order >= MAX_ORDER)
938                 return;
939
940         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
941
942         /*
943          * We want to release as many surplus pages as possible, spread
944          * evenly across all nodes. Iterate across all nodes until we
945          * can no longer free unreserved surplus pages. This occurs when
946          * the nodes with surplus pages have no free pages.
947          * free_pool_huge_page() will balance the the frees across the
948          * on-line nodes for us and will handle the hstate accounting.
949          */
950         while (nr_pages--) {
951                 if (!free_pool_huge_page(h, &node_online_map, 1))
952                         break;
953         }
954 }
955
956 /*
957  * Determine if the huge page at addr within the vma has an associated
958  * reservation.  Where it does not we will need to logically increase
959  * reservation and actually increase quota before an allocation can occur.
960  * Where any new reservation would be required the reservation change is
961  * prepared, but not committed.  Once the page has been quota'd allocated
962  * an instantiated the change should be committed via vma_commit_reservation.
963  * No action is required on failure.
964  */
965 static long vma_needs_reservation(struct hstate *h,
966                         struct vm_area_struct *vma, unsigned long addr)
967 {
968         struct address_space *mapping = vma->vm_file->f_mapping;
969         struct inode *inode = mapping->host;
970
971         if (vma->vm_flags & VM_MAYSHARE) {
972                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
973                 return region_chg(&inode->i_mapping->private_list,
974                                                         idx, idx + 1);
975
976         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
977                 return 1;
978
979         } else  {
980                 long err;
981                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
982                 struct resv_map *reservations = vma_resv_map(vma);
983
984                 err = region_chg(&reservations->regions, idx, idx + 1);
985                 if (err < 0)
986                         return err;
987                 return 0;
988         }
989 }
990 static void vma_commit_reservation(struct hstate *h,
991                         struct vm_area_struct *vma, unsigned long addr)
992 {
993         struct address_space *mapping = vma->vm_file->f_mapping;
994         struct inode *inode = mapping->host;
995
996         if (vma->vm_flags & VM_MAYSHARE) {
997                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
998                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
999
1000         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1001                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1002                 struct resv_map *reservations = vma_resv_map(vma);
1003
1004                 /* Mark this page used in the map. */
1005                 region_add(&reservations->regions, idx, idx + 1);
1006         }
1007 }
1008
1009 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1010                                     unsigned long addr, int avoid_reserve)
1011 {
1012         struct hstate *h = hstate_vma(vma);
1013         struct page *page;
1014         struct address_space *mapping = vma->vm_file->f_mapping;
1015         struct inode *inode = mapping->host;
1016         long chg;
1017
1018         /*
1019          * Processes that did not create the mapping will have no reserves and
1020          * will not have accounted against quota. Check that the quota can be
1021          * made before satisfying the allocation
1022          * MAP_NORESERVE mappings may also need pages and quota allocated
1023          * if no reserve mapping overlaps.
1024          */
1025         chg = vma_needs_reservation(h, vma, addr);
1026         if (chg < 0)
1027                 return ERR_PTR(chg);
1028         if (chg)
1029                 if (hugetlb_get_quota(inode->i_mapping, chg))
1030                         return ERR_PTR(-ENOSPC);
1031
1032         spin_lock(&hugetlb_lock);
1033         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1034         spin_unlock(&hugetlb_lock);
1035
1036         if (!page) {
1037                 page = alloc_buddy_huge_page(h, vma, addr);
1038                 if (!page) {
1039                         hugetlb_put_quota(inode->i_mapping, chg);
1040                         return ERR_PTR(-VM_FAULT_OOM);
1041                 }
1042         }
1043
1044         set_page_refcounted(page);
1045         set_page_private(page, (unsigned long) mapping);
1046
1047         vma_commit_reservation(h, vma, addr);
1048
1049         return page;
1050 }
1051
1052 int __weak alloc_bootmem_huge_page(struct hstate *h)
1053 {
1054         struct huge_bootmem_page *m;
1055         int nr_nodes = nodes_weight(node_online_map);
1056
1057         while (nr_nodes) {
1058                 void *addr;
1059
1060                 addr = __alloc_bootmem_node_nopanic(
1061                                 NODE_DATA(hstate_next_node_to_alloc(h,
1062                                                         &node_online_map)),
1063                                 huge_page_size(h), huge_page_size(h), 0);
1064
1065                 if (addr) {
1066                         /*
1067                          * Use the beginning of the huge page to store the
1068                          * huge_bootmem_page struct (until gather_bootmem
1069                          * puts them into the mem_map).
1070                          */
1071                         m = addr;
1072                         goto found;
1073                 }
1074                 nr_nodes--;
1075         }
1076         return 0;
1077
1078 found:
1079         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1080         /* Put them into a private list first because mem_map is not up yet */
1081         list_add(&m->list, &huge_boot_pages);
1082         m->hstate = h;
1083         return 1;
1084 }
1085
1086 static void prep_compound_huge_page(struct page *page, int order)
1087 {
1088         if (unlikely(order > (MAX_ORDER - 1)))
1089                 prep_compound_gigantic_page(page, order);
1090         else
1091                 prep_compound_page(page, order);
1092 }
1093
1094 /* Put bootmem huge pages into the standard lists after mem_map is up */
1095 static void __init gather_bootmem_prealloc(void)
1096 {
1097         struct huge_bootmem_page *m;
1098
1099         list_for_each_entry(m, &huge_boot_pages, list) {
1100                 struct page *page = virt_to_page(m);
1101                 struct hstate *h = m->hstate;
1102                 __ClearPageReserved(page);
1103                 WARN_ON(page_count(page) != 1);
1104                 prep_compound_huge_page(page, h->order);
1105                 prep_new_huge_page(h, page, page_to_nid(page));
1106         }
1107 }
1108
1109 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1110 {
1111         unsigned long i;
1112
1113         for (i = 0; i < h->max_huge_pages; ++i) {
1114                 if (h->order >= MAX_ORDER) {
1115                         if (!alloc_bootmem_huge_page(h))
1116                                 break;
1117                 } else if (!alloc_fresh_huge_page(h, &node_online_map))
1118                         break;
1119         }
1120         h->max_huge_pages = i;
1121 }
1122
1123 static void __init hugetlb_init_hstates(void)
1124 {
1125         struct hstate *h;
1126
1127         for_each_hstate(h) {
1128                 /* oversize hugepages were init'ed in early boot */
1129                 if (h->order < MAX_ORDER)
1130                         hugetlb_hstate_alloc_pages(h);
1131         }
1132 }
1133
1134 static char * __init memfmt(char *buf, unsigned long n)
1135 {
1136         if (n >= (1UL << 30))
1137                 sprintf(buf, "%lu GB", n >> 30);
1138         else if (n >= (1UL << 20))
1139                 sprintf(buf, "%lu MB", n >> 20);
1140         else
1141                 sprintf(buf, "%lu KB", n >> 10);
1142         return buf;
1143 }
1144
1145 static void __init report_hugepages(void)
1146 {
1147         struct hstate *h;
1148
1149         for_each_hstate(h) {
1150                 char buf[32];
1151                 printk(KERN_INFO "HugeTLB registered %s page size, "
1152                                  "pre-allocated %ld pages\n",
1153                         memfmt(buf, huge_page_size(h)),
1154                         h->free_huge_pages);
1155         }
1156 }
1157
1158 #ifdef CONFIG_HIGHMEM
1159 static void try_to_free_low(struct hstate *h, unsigned long count,
1160                                                 nodemask_t *nodes_allowed)
1161 {
1162         int i;
1163
1164         if (h->order >= MAX_ORDER)
1165                 return;
1166
1167         for_each_node_mask(i, *nodes_allowed) {
1168                 struct page *page, *next;
1169                 struct list_head *freel = &h->hugepage_freelists[i];
1170                 list_for_each_entry_safe(page, next, freel, lru) {
1171                         if (count >= h->nr_huge_pages)
1172                                 return;
1173                         if (PageHighMem(page))
1174                                 continue;
1175                         list_del(&page->lru);
1176                         update_and_free_page(h, page);
1177                         h->free_huge_pages--;
1178                         h->free_huge_pages_node[page_to_nid(page)]--;
1179                 }
1180         }
1181 }
1182 #else
1183 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1184                                                 nodemask_t *nodes_allowed)
1185 {
1186 }
1187 #endif
1188
1189 /*
1190  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1191  * balanced by operating on them in a round-robin fashion.
1192  * Returns 1 if an adjustment was made.
1193  */
1194 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1195                                 int delta)
1196 {
1197         int start_nid, next_nid;
1198         int ret = 0;
1199
1200         VM_BUG_ON(delta != -1 && delta != 1);
1201
1202         if (delta < 0)
1203                 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1204         else
1205                 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1206         next_nid = start_nid;
1207
1208         do {
1209                 int nid = next_nid;
1210                 if (delta < 0)  {
1211                         /*
1212                          * To shrink on this node, there must be a surplus page
1213                          */
1214                         if (!h->surplus_huge_pages_node[nid]) {
1215                                 next_nid = hstate_next_node_to_alloc(h,
1216                                                                 nodes_allowed);
1217                                 continue;
1218                         }
1219                 }
1220                 if (delta > 0) {
1221                         /*
1222                          * Surplus cannot exceed the total number of pages
1223                          */
1224                         if (h->surplus_huge_pages_node[nid] >=
1225                                                 h->nr_huge_pages_node[nid]) {
1226                                 next_nid = hstate_next_node_to_free(h,
1227                                                                 nodes_allowed);
1228                                 continue;
1229                         }
1230                 }
1231
1232                 h->surplus_huge_pages += delta;
1233                 h->surplus_huge_pages_node[nid] += delta;
1234                 ret = 1;
1235                 break;
1236         } while (next_nid != start_nid);
1237
1238         return ret;
1239 }
1240
1241 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1242 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1243                                                 nodemask_t *nodes_allowed)
1244 {
1245         unsigned long min_count, ret;
1246
1247         if (h->order >= MAX_ORDER)
1248                 return h->max_huge_pages;
1249
1250         /*
1251          * Increase the pool size
1252          * First take pages out of surplus state.  Then make up the
1253          * remaining difference by allocating fresh huge pages.
1254          *
1255          * We might race with alloc_buddy_huge_page() here and be unable
1256          * to convert a surplus huge page to a normal huge page. That is
1257          * not critical, though, it just means the overall size of the
1258          * pool might be one hugepage larger than it needs to be, but
1259          * within all the constraints specified by the sysctls.
1260          */
1261         spin_lock(&hugetlb_lock);
1262         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1263                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1264                         break;
1265         }
1266
1267         while (count > persistent_huge_pages(h)) {
1268                 /*
1269                  * If this allocation races such that we no longer need the
1270                  * page, free_huge_page will handle it by freeing the page
1271                  * and reducing the surplus.
1272                  */
1273                 spin_unlock(&hugetlb_lock);
1274                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1275                 spin_lock(&hugetlb_lock);
1276                 if (!ret)
1277                         goto out;
1278
1279         }
1280
1281         /*
1282          * Decrease the pool size
1283          * First return free pages to the buddy allocator (being careful
1284          * to keep enough around to satisfy reservations).  Then place
1285          * pages into surplus state as needed so the pool will shrink
1286          * to the desired size as pages become free.
1287          *
1288          * By placing pages into the surplus state independent of the
1289          * overcommit value, we are allowing the surplus pool size to
1290          * exceed overcommit. There are few sane options here. Since
1291          * alloc_buddy_huge_page() is checking the global counter,
1292          * though, we'll note that we're not allowed to exceed surplus
1293          * and won't grow the pool anywhere else. Not until one of the
1294          * sysctls are changed, or the surplus pages go out of use.
1295          */
1296         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1297         min_count = max(count, min_count);
1298         try_to_free_low(h, min_count, nodes_allowed);
1299         while (min_count < persistent_huge_pages(h)) {
1300                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1301                         break;
1302         }
1303         while (count < persistent_huge_pages(h)) {
1304                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1305                         break;
1306         }
1307 out:
1308         ret = persistent_huge_pages(h);
1309         spin_unlock(&hugetlb_lock);
1310         return ret;
1311 }
1312
1313 #define HSTATE_ATTR_RO(_name) \
1314         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1315
1316 #define HSTATE_ATTR(_name) \
1317         static struct kobj_attribute _name##_attr = \
1318                 __ATTR(_name, 0644, _name##_show, _name##_store)
1319
1320 static struct kobject *hugepages_kobj;
1321 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1322
1323 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1324 {
1325         int i;
1326         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1327                 if (hstate_kobjs[i] == kobj)
1328                         return &hstates[i];
1329         BUG();
1330         return NULL;
1331 }
1332
1333 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1334                                         struct kobj_attribute *attr, char *buf)
1335 {
1336         struct hstate *h = kobj_to_hstate(kobj);
1337         return sprintf(buf, "%lu\n", h->nr_huge_pages);
1338 }
1339 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1340                         struct kobject *kobj, struct kobj_attribute *attr,
1341                         const char *buf, size_t len)
1342 {
1343         int err;
1344         unsigned long count;
1345         struct hstate *h = kobj_to_hstate(kobj);
1346         NODEMASK_ALLOC(nodemask_t, nodes_allowed);
1347
1348         err = strict_strtoul(buf, 10, &count);
1349         if (err)
1350                 return 0;
1351
1352         if (!(obey_mempolicy && init_nodemask_of_mempolicy(nodes_allowed))) {
1353                 NODEMASK_FREE(nodes_allowed);
1354                 nodes_allowed = &node_online_map;
1355         }
1356         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1357
1358         if (nodes_allowed != &node_online_map)
1359                 NODEMASK_FREE(nodes_allowed);
1360
1361         return len;
1362 }
1363
1364 static ssize_t nr_hugepages_show(struct kobject *kobj,
1365                                        struct kobj_attribute *attr, char *buf)
1366 {
1367         return nr_hugepages_show_common(kobj, attr, buf);
1368 }
1369
1370 static ssize_t nr_hugepages_store(struct kobject *kobj,
1371                struct kobj_attribute *attr, const char *buf, size_t len)
1372 {
1373         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1374 }
1375 HSTATE_ATTR(nr_hugepages);
1376
1377 #ifdef CONFIG_NUMA
1378
1379 /*
1380  * hstate attribute for optionally mempolicy-based constraint on persistent
1381  * huge page alloc/free.
1382  */
1383 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1384                                        struct kobj_attribute *attr, char *buf)
1385 {
1386         return nr_hugepages_show_common(kobj, attr, buf);
1387 }
1388
1389 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1390                struct kobj_attribute *attr, const char *buf, size_t len)
1391 {
1392         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1393 }
1394 HSTATE_ATTR(nr_hugepages_mempolicy);
1395 #endif
1396
1397
1398 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1399                                         struct kobj_attribute *attr, char *buf)
1400 {
1401         struct hstate *h = kobj_to_hstate(kobj);
1402         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1403 }
1404 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1405                 struct kobj_attribute *attr, const char *buf, size_t count)
1406 {
1407         int err;
1408         unsigned long input;
1409         struct hstate *h = kobj_to_hstate(kobj);
1410
1411         err = strict_strtoul(buf, 10, &input);
1412         if (err)
1413                 return 0;
1414
1415         spin_lock(&hugetlb_lock);
1416         h->nr_overcommit_huge_pages = input;
1417         spin_unlock(&hugetlb_lock);
1418
1419         return count;
1420 }
1421 HSTATE_ATTR(nr_overcommit_hugepages);
1422
1423 static ssize_t free_hugepages_show(struct kobject *kobj,
1424                                         struct kobj_attribute *attr, char *buf)
1425 {
1426         struct hstate *h = kobj_to_hstate(kobj);
1427         return sprintf(buf, "%lu\n", h->free_huge_pages);
1428 }
1429 HSTATE_ATTR_RO(free_hugepages);
1430
1431 static ssize_t resv_hugepages_show(struct kobject *kobj,
1432                                         struct kobj_attribute *attr, char *buf)
1433 {
1434         struct hstate *h = kobj_to_hstate(kobj);
1435         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1436 }
1437 HSTATE_ATTR_RO(resv_hugepages);
1438
1439 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1440                                         struct kobj_attribute *attr, char *buf)
1441 {
1442         struct hstate *h = kobj_to_hstate(kobj);
1443         return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1444 }
1445 HSTATE_ATTR_RO(surplus_hugepages);
1446
1447 static struct attribute *hstate_attrs[] = {
1448         &nr_hugepages_attr.attr,
1449         &nr_overcommit_hugepages_attr.attr,
1450         &free_hugepages_attr.attr,
1451         &resv_hugepages_attr.attr,
1452         &surplus_hugepages_attr.attr,
1453 #ifdef CONFIG_NUMA
1454         &nr_hugepages_mempolicy_attr.attr,
1455 #endif
1456         NULL,
1457 };
1458
1459 static struct attribute_group hstate_attr_group = {
1460         .attrs = hstate_attrs,
1461 };
1462
1463 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1464 {
1465         int retval;
1466
1467         hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1468                                                         hugepages_kobj);
1469         if (!hstate_kobjs[h - hstates])
1470                 return -ENOMEM;
1471
1472         retval = sysfs_create_group(hstate_kobjs[h - hstates],
1473                                                         &hstate_attr_group);
1474         if (retval)
1475                 kobject_put(hstate_kobjs[h - hstates]);
1476
1477         return retval;
1478 }
1479
1480 static void __init hugetlb_sysfs_init(void)
1481 {
1482         struct hstate *h;
1483         int err;
1484
1485         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1486         if (!hugepages_kobj)
1487                 return;
1488
1489         for_each_hstate(h) {
1490                 err = hugetlb_sysfs_add_hstate(h);
1491                 if (err)
1492                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1493                                                                 h->name);
1494         }
1495 }
1496
1497 static void __exit hugetlb_exit(void)
1498 {
1499         struct hstate *h;
1500
1501         for_each_hstate(h) {
1502                 kobject_put(hstate_kobjs[h - hstates]);
1503         }
1504
1505         kobject_put(hugepages_kobj);
1506 }
1507 module_exit(hugetlb_exit);
1508
1509 static int __init hugetlb_init(void)
1510 {
1511         /* Some platform decide whether they support huge pages at boot
1512          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1513          * there is no such support
1514          */
1515         if (HPAGE_SHIFT == 0)
1516                 return 0;
1517
1518         if (!size_to_hstate(default_hstate_size)) {
1519                 default_hstate_size = HPAGE_SIZE;
1520                 if (!size_to_hstate(default_hstate_size))
1521                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1522         }
1523         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1524         if (default_hstate_max_huge_pages)
1525                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1526
1527         hugetlb_init_hstates();
1528
1529         gather_bootmem_prealloc();
1530
1531         report_hugepages();
1532
1533         hugetlb_sysfs_init();
1534
1535         return 0;
1536 }
1537 module_init(hugetlb_init);
1538
1539 /* Should be called on processing a hugepagesz=... option */
1540 void __init hugetlb_add_hstate(unsigned order)
1541 {
1542         struct hstate *h;
1543         unsigned long i;
1544
1545         if (size_to_hstate(PAGE_SIZE << order)) {
1546                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1547                 return;
1548         }
1549         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1550         BUG_ON(order == 0);
1551         h = &hstates[max_hstate++];
1552         h->order = order;
1553         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1554         h->nr_huge_pages = 0;
1555         h->free_huge_pages = 0;
1556         for (i = 0; i < MAX_NUMNODES; ++i)
1557                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1558         h->next_nid_to_alloc = first_node(node_online_map);
1559         h->next_nid_to_free = first_node(node_online_map);
1560         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1561                                         huge_page_size(h)/1024);
1562
1563         parsed_hstate = h;
1564 }
1565
1566 static int __init hugetlb_nrpages_setup(char *s)
1567 {
1568         unsigned long *mhp;
1569         static unsigned long *last_mhp;
1570
1571         /*
1572          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1573          * so this hugepages= parameter goes to the "default hstate".
1574          */
1575         if (!max_hstate)
1576                 mhp = &default_hstate_max_huge_pages;
1577         else
1578                 mhp = &parsed_hstate->max_huge_pages;
1579
1580         if (mhp == last_mhp) {
1581                 printk(KERN_WARNING "hugepages= specified twice without "
1582                         "interleaving hugepagesz=, ignoring\n");
1583                 return 1;
1584         }
1585
1586         if (sscanf(s, "%lu", mhp) <= 0)
1587                 *mhp = 0;
1588
1589         /*
1590          * Global state is always initialized later in hugetlb_init.
1591          * But we need to allocate >= MAX_ORDER hstates here early to still
1592          * use the bootmem allocator.
1593          */
1594         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1595                 hugetlb_hstate_alloc_pages(parsed_hstate);
1596
1597         last_mhp = mhp;
1598
1599         return 1;
1600 }
1601 __setup("hugepages=", hugetlb_nrpages_setup);
1602
1603 static int __init hugetlb_default_setup(char *s)
1604 {
1605         default_hstate_size = memparse(s, &s);
1606         return 1;
1607 }
1608 __setup("default_hugepagesz=", hugetlb_default_setup);
1609
1610 static unsigned int cpuset_mems_nr(unsigned int *array)
1611 {
1612         int node;
1613         unsigned int nr = 0;
1614
1615         for_each_node_mask(node, cpuset_current_mems_allowed)
1616                 nr += array[node];
1617
1618         return nr;
1619 }
1620
1621 #ifdef CONFIG_SYSCTL
1622 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1623                          struct ctl_table *table, int write,
1624                          void __user *buffer, size_t *length, loff_t *ppos)
1625 {
1626         struct hstate *h = &default_hstate;
1627         unsigned long tmp;
1628
1629         if (!write)
1630                 tmp = h->max_huge_pages;
1631
1632         table->data = &tmp;
1633         table->maxlen = sizeof(unsigned long);
1634         proc_doulongvec_minmax(table, write, buffer, length, ppos);
1635
1636         if (write) {
1637                 NODEMASK_ALLOC(nodemask_t, nodes_allowed);
1638                 if (!(obey_mempolicy &&
1639                                init_nodemask_of_mempolicy(nodes_allowed))) {
1640                         NODEMASK_FREE(nodes_allowed);
1641                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1642                 }
1643                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1644
1645                 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1646                         NODEMASK_FREE(nodes_allowed);
1647         }
1648
1649         return 0;
1650 }
1651
1652 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1653                           void __user *buffer, size_t *length, loff_t *ppos)
1654 {
1655
1656         return hugetlb_sysctl_handler_common(false, table, write,
1657                                                         buffer, length, ppos);
1658 }
1659
1660 #ifdef CONFIG_NUMA
1661 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1662                           void __user *buffer, size_t *length, loff_t *ppos)
1663 {
1664         return hugetlb_sysctl_handler_common(true, table, write,
1665                                                         buffer, length, ppos);
1666 }
1667 #endif /* CONFIG_NUMA */
1668
1669 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1670                         void __user *buffer,
1671                         size_t *length, loff_t *ppos)
1672 {
1673         proc_dointvec(table, write, buffer, length, ppos);
1674         if (hugepages_treat_as_movable)
1675                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1676         else
1677                 htlb_alloc_mask = GFP_HIGHUSER;
1678         return 0;
1679 }
1680
1681 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1682                         void __user *buffer,
1683                         size_t *length, loff_t *ppos)
1684 {
1685         struct hstate *h = &default_hstate;
1686         unsigned long tmp;
1687
1688         if (!write)
1689                 tmp = h->nr_overcommit_huge_pages;
1690
1691         table->data = &tmp;
1692         table->maxlen = sizeof(unsigned long);
1693         proc_doulongvec_minmax(table, write, buffer, length, ppos);
1694
1695         if (write) {
1696                 spin_lock(&hugetlb_lock);
1697                 h->nr_overcommit_huge_pages = tmp;
1698                 spin_unlock(&hugetlb_lock);
1699         }
1700
1701         return 0;
1702 }
1703
1704 #endif /* CONFIG_SYSCTL */
1705
1706 void hugetlb_report_meminfo(struct seq_file *m)
1707 {
1708         struct hstate *h = &default_hstate;
1709         seq_printf(m,
1710                         "HugePages_Total:   %5lu\n"
1711                         "HugePages_Free:    %5lu\n"
1712                         "HugePages_Rsvd:    %5lu\n"
1713                         "HugePages_Surp:    %5lu\n"
1714                         "Hugepagesize:   %8lu kB\n",
1715                         h->nr_huge_pages,
1716                         h->free_huge_pages,
1717                         h->resv_huge_pages,
1718                         h->surplus_huge_pages,
1719                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1720 }
1721
1722 int hugetlb_report_node_meminfo(int nid, char *buf)
1723 {
1724         struct hstate *h = &default_hstate;
1725         return sprintf(buf,
1726                 "Node %d HugePages_Total: %5u\n"
1727                 "Node %d HugePages_Free:  %5u\n"
1728                 "Node %d HugePages_Surp:  %5u\n",
1729                 nid, h->nr_huge_pages_node[nid],
1730                 nid, h->free_huge_pages_node[nid],
1731                 nid, h->surplus_huge_pages_node[nid]);
1732 }
1733
1734 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1735 unsigned long hugetlb_total_pages(void)
1736 {
1737         struct hstate *h = &default_hstate;
1738         return h->nr_huge_pages * pages_per_huge_page(h);
1739 }
1740
1741 static int hugetlb_acct_memory(struct hstate *h, long delta)
1742 {
1743         int ret = -ENOMEM;
1744
1745         spin_lock(&hugetlb_lock);
1746         /*
1747          * When cpuset is configured, it breaks the strict hugetlb page
1748          * reservation as the accounting is done on a global variable. Such
1749          * reservation is completely rubbish in the presence of cpuset because
1750          * the reservation is not checked against page availability for the
1751          * current cpuset. Application can still potentially OOM'ed by kernel
1752          * with lack of free htlb page in cpuset that the task is in.
1753          * Attempt to enforce strict accounting with cpuset is almost
1754          * impossible (or too ugly) because cpuset is too fluid that
1755          * task or memory node can be dynamically moved between cpusets.
1756          *
1757          * The change of semantics for shared hugetlb mapping with cpuset is
1758          * undesirable. However, in order to preserve some of the semantics,
1759          * we fall back to check against current free page availability as
1760          * a best attempt and hopefully to minimize the impact of changing
1761          * semantics that cpuset has.
1762          */
1763         if (delta > 0) {
1764                 if (gather_surplus_pages(h, delta) < 0)
1765                         goto out;
1766
1767                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1768                         return_unused_surplus_pages(h, delta);
1769                         goto out;
1770                 }
1771         }
1772
1773         ret = 0;
1774         if (delta < 0)
1775                 return_unused_surplus_pages(h, (unsigned long) -delta);
1776
1777 out:
1778         spin_unlock(&hugetlb_lock);
1779         return ret;
1780 }
1781
1782 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1783 {
1784         struct resv_map *reservations = vma_resv_map(vma);
1785
1786         /*
1787          * This new VMA should share its siblings reservation map if present.
1788          * The VMA will only ever have a valid reservation map pointer where
1789          * it is being copied for another still existing VMA.  As that VMA
1790          * has a reference to the reservation map it cannot dissappear until
1791          * after this open call completes.  It is therefore safe to take a
1792          * new reference here without additional locking.
1793          */
1794         if (reservations)
1795                 kref_get(&reservations->refs);
1796 }
1797
1798 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1799 {
1800         struct hstate *h = hstate_vma(vma);
1801         struct resv_map *reservations = vma_resv_map(vma);
1802         unsigned long reserve;
1803         unsigned long start;
1804         unsigned long end;
1805
1806         if (reservations) {
1807                 start = vma_hugecache_offset(h, vma, vma->vm_start);
1808                 end = vma_hugecache_offset(h, vma, vma->vm_end);
1809
1810                 reserve = (end - start) -
1811                         region_count(&reservations->regions, start, end);
1812
1813                 kref_put(&reservations->refs, resv_map_release);
1814
1815                 if (reserve) {
1816                         hugetlb_acct_memory(h, -reserve);
1817                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1818                 }
1819         }
1820 }
1821
1822 /*
1823  * We cannot handle pagefaults against hugetlb pages at all.  They cause
1824  * handle_mm_fault() to try to instantiate regular-sized pages in the
1825  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
1826  * this far.
1827  */
1828 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1829 {
1830         BUG();
1831         return 0;
1832 }
1833
1834 const struct vm_operations_struct hugetlb_vm_ops = {
1835         .fault = hugetlb_vm_op_fault,
1836         .open = hugetlb_vm_op_open,
1837         .close = hugetlb_vm_op_close,
1838 };
1839
1840 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1841                                 int writable)
1842 {
1843         pte_t entry;
1844
1845         if (writable) {
1846                 entry =
1847                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1848         } else {
1849                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1850         }
1851         entry = pte_mkyoung(entry);
1852         entry = pte_mkhuge(entry);
1853
1854         return entry;
1855 }
1856
1857 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1858                                    unsigned long address, pte_t *ptep)
1859 {
1860         pte_t entry;
1861
1862         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1863         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1864                 update_mmu_cache(vma, address, entry);
1865         }
1866 }
1867
1868
1869 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1870                             struct vm_area_struct *vma)
1871 {
1872         pte_t *src_pte, *dst_pte, entry;
1873         struct page *ptepage;
1874         unsigned long addr;
1875         int cow;
1876         struct hstate *h = hstate_vma(vma);
1877         unsigned long sz = huge_page_size(h);
1878
1879         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1880
1881         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1882                 src_pte = huge_pte_offset(src, addr);
1883                 if (!src_pte)
1884                         continue;
1885                 dst_pte = huge_pte_alloc(dst, addr, sz);
1886                 if (!dst_pte)
1887                         goto nomem;
1888
1889                 /* If the pagetables are shared don't copy or take references */
1890                 if (dst_pte == src_pte)
1891                         continue;
1892
1893                 spin_lock(&dst->page_table_lock);
1894                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1895                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1896                         if (cow)
1897                                 huge_ptep_set_wrprotect(src, addr, src_pte);
1898                         entry = huge_ptep_get(src_pte);
1899                         ptepage = pte_page(entry);
1900                         get_page(ptepage);
1901                         set_huge_pte_at(dst, addr, dst_pte, entry);
1902                 }
1903                 spin_unlock(&src->page_table_lock);
1904                 spin_unlock(&dst->page_table_lock);
1905         }
1906         return 0;
1907
1908 nomem:
1909         return -ENOMEM;
1910 }
1911
1912 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1913                             unsigned long end, struct page *ref_page)
1914 {
1915         struct mm_struct *mm = vma->vm_mm;
1916         unsigned long address;
1917         pte_t *ptep;
1918         pte_t pte;
1919         struct page *page;
1920         struct page *tmp;
1921         struct hstate *h = hstate_vma(vma);
1922         unsigned long sz = huge_page_size(h);
1923
1924         /*
1925          * A page gathering list, protected by per file i_mmap_lock. The
1926          * lock is used to avoid list corruption from multiple unmapping
1927          * of the same page since we are using page->lru.
1928          */
1929         LIST_HEAD(page_list);
1930
1931         WARN_ON(!is_vm_hugetlb_page(vma));
1932         BUG_ON(start & ~huge_page_mask(h));
1933         BUG_ON(end & ~huge_page_mask(h));
1934
1935         mmu_notifier_invalidate_range_start(mm, start, end);
1936         spin_lock(&mm->page_table_lock);
1937         for (address = start; address < end; address += sz) {
1938                 ptep = huge_pte_offset(mm, address);
1939                 if (!ptep)
1940                         continue;
1941
1942                 if (huge_pmd_unshare(mm, &address, ptep))
1943                         continue;
1944
1945                 /*
1946                  * If a reference page is supplied, it is because a specific
1947                  * page is being unmapped, not a range. Ensure the page we
1948                  * are about to unmap is the actual page of interest.
1949                  */
1950                 if (ref_page) {
1951                         pte = huge_ptep_get(ptep);
1952                         if (huge_pte_none(pte))
1953                                 continue;
1954                         page = pte_page(pte);
1955                         if (page != ref_page)
1956                                 continue;
1957
1958                         /*
1959                          * Mark the VMA as having unmapped its page so that
1960                          * future faults in this VMA will fail rather than
1961                          * looking like data was lost
1962                          */
1963                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1964                 }
1965
1966                 pte = huge_ptep_get_and_clear(mm, address, ptep);
1967                 if (huge_pte_none(pte))
1968                         continue;
1969
1970                 page = pte_page(pte);
1971                 if (pte_dirty(pte))
1972                         set_page_dirty(page);
1973                 list_add(&page->lru, &page_list);
1974         }
1975         spin_unlock(&mm->page_table_lock);
1976         flush_tlb_range(vma, start, end);
1977         mmu_notifier_invalidate_range_end(mm, start, end);
1978         list_for_each_entry_safe(page, tmp, &page_list, lru) {
1979                 list_del(&page->lru);
1980                 put_page(page);
1981         }
1982 }
1983
1984 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1985                           unsigned long end, struct page *ref_page)
1986 {
1987         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1988         __unmap_hugepage_range(vma, start, end, ref_page);
1989         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1990 }
1991
1992 /*
1993  * This is called when the original mapper is failing to COW a MAP_PRIVATE
1994  * mappping it owns the reserve page for. The intention is to unmap the page
1995  * from other VMAs and let the children be SIGKILLed if they are faulting the
1996  * same region.
1997  */
1998 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1999                                 struct page *page, unsigned long address)
2000 {
2001         struct hstate *h = hstate_vma(vma);
2002         struct vm_area_struct *iter_vma;
2003         struct address_space *mapping;
2004         struct prio_tree_iter iter;
2005         pgoff_t pgoff;
2006
2007         /*
2008          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2009          * from page cache lookup which is in HPAGE_SIZE units.
2010          */
2011         address = address & huge_page_mask(h);
2012         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2013                 + (vma->vm_pgoff >> PAGE_SHIFT);
2014         mapping = (struct address_space *)page_private(page);
2015
2016         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2017                 /* Do not unmap the current VMA */
2018                 if (iter_vma == vma)
2019                         continue;
2020
2021                 /*
2022                  * Unmap the page from other VMAs without their own reserves.
2023                  * They get marked to be SIGKILLed if they fault in these
2024                  * areas. This is because a future no-page fault on this VMA
2025                  * could insert a zeroed page instead of the data existing
2026                  * from the time of fork. This would look like data corruption
2027                  */
2028                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2029                         unmap_hugepage_range(iter_vma,
2030                                 address, address + huge_page_size(h),
2031                                 page);
2032         }
2033
2034         return 1;
2035 }
2036
2037 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2038                         unsigned long address, pte_t *ptep, pte_t pte,
2039                         struct page *pagecache_page)
2040 {
2041         struct hstate *h = hstate_vma(vma);
2042         struct page *old_page, *new_page;
2043         int avoidcopy;
2044         int outside_reserve = 0;
2045
2046         old_page = pte_page(pte);
2047
2048 retry_avoidcopy:
2049         /* If no-one else is actually using this page, avoid the copy
2050          * and just make the page writable */
2051         avoidcopy = (page_count(old_page) == 1);
2052         if (avoidcopy) {
2053                 set_huge_ptep_writable(vma, address, ptep);
2054                 return 0;
2055         }
2056
2057         /*
2058          * If the process that created a MAP_PRIVATE mapping is about to
2059          * perform a COW due to a shared page count, attempt to satisfy
2060          * the allocation without using the existing reserves. The pagecache
2061          * page is used to determine if the reserve at this address was
2062          * consumed or not. If reserves were used, a partial faulted mapping
2063          * at the time of fork() could consume its reserves on COW instead
2064          * of the full address range.
2065          */
2066         if (!(vma->vm_flags & VM_MAYSHARE) &&
2067                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2068                         old_page != pagecache_page)
2069                 outside_reserve = 1;
2070
2071         page_cache_get(old_page);
2072         new_page = alloc_huge_page(vma, address, outside_reserve);
2073
2074         if (IS_ERR(new_page)) {
2075                 page_cache_release(old_page);
2076
2077                 /*
2078                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2079                  * it is due to references held by a child and an insufficient
2080                  * huge page pool. To guarantee the original mappers
2081                  * reliability, unmap the page from child processes. The child
2082                  * may get SIGKILLed if it later faults.
2083                  */
2084                 if (outside_reserve) {
2085                         BUG_ON(huge_pte_none(pte));
2086                         if (unmap_ref_private(mm, vma, old_page, address)) {
2087                                 BUG_ON(page_count(old_page) != 1);
2088                                 BUG_ON(huge_pte_none(pte));
2089                                 goto retry_avoidcopy;
2090                         }
2091                         WARN_ON_ONCE(1);
2092                 }
2093
2094                 return -PTR_ERR(new_page);
2095         }
2096
2097         spin_unlock(&mm->page_table_lock);
2098         copy_huge_page(new_page, old_page, address, vma);
2099         __SetPageUptodate(new_page);
2100         spin_lock(&mm->page_table_lock);
2101
2102         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2103         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2104                 /* Break COW */
2105                 huge_ptep_clear_flush(vma, address, ptep);
2106                 set_huge_pte_at(mm, address, ptep,
2107                                 make_huge_pte(vma, new_page, 1));
2108                 /* Make the old page be freed below */
2109                 new_page = old_page;
2110         }
2111         page_cache_release(new_page);
2112         page_cache_release(old_page);
2113         return 0;
2114 }
2115
2116 /* Return the pagecache page at a given address within a VMA */
2117 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2118                         struct vm_area_struct *vma, unsigned long address)
2119 {
2120         struct address_space *mapping;
2121         pgoff_t idx;
2122
2123         mapping = vma->vm_file->f_mapping;
2124         idx = vma_hugecache_offset(h, vma, address);
2125
2126         return find_lock_page(mapping, idx);
2127 }
2128
2129 /*
2130  * Return whether there is a pagecache page to back given address within VMA.
2131  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2132  */
2133 static bool hugetlbfs_pagecache_present(struct hstate *h,
2134                         struct vm_area_struct *vma, unsigned long address)
2135 {
2136         struct address_space *mapping;
2137         pgoff_t idx;
2138         struct page *page;
2139
2140         mapping = vma->vm_file->f_mapping;
2141         idx = vma_hugecache_offset(h, vma, address);
2142
2143         page = find_get_page(mapping, idx);
2144         if (page)
2145                 put_page(page);
2146         return page != NULL;
2147 }
2148
2149 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2150                         unsigned long address, pte_t *ptep, unsigned int flags)
2151 {
2152         struct hstate *h = hstate_vma(vma);
2153         int ret = VM_FAULT_SIGBUS;
2154         pgoff_t idx;
2155         unsigned long size;
2156         struct page *page;
2157         struct address_space *mapping;
2158         pte_t new_pte;
2159
2160         /*
2161          * Currently, we are forced to kill the process in the event the
2162          * original mapper has unmapped pages from the child due to a failed
2163          * COW. Warn that such a situation has occured as it may not be obvious
2164          */
2165         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2166                 printk(KERN_WARNING
2167                         "PID %d killed due to inadequate hugepage pool\n",
2168                         current->pid);
2169                 return ret;
2170         }
2171
2172         mapping = vma->vm_file->f_mapping;
2173         idx = vma_hugecache_offset(h, vma, address);
2174
2175         /*
2176          * Use page lock to guard against racing truncation
2177          * before we get page_table_lock.
2178          */
2179 retry:
2180         page = find_lock_page(mapping, idx);
2181         if (!page) {
2182                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2183                 if (idx >= size)
2184                         goto out;
2185                 page = alloc_huge_page(vma, address, 0);
2186                 if (IS_ERR(page)) {
2187                         ret = -PTR_ERR(page);
2188                         goto out;
2189                 }
2190                 clear_huge_page(page, address, huge_page_size(h));
2191                 __SetPageUptodate(page);
2192
2193                 if (vma->vm_flags & VM_MAYSHARE) {
2194                         int err;
2195                         struct inode *inode = mapping->host;
2196
2197                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2198                         if (err) {
2199                                 put_page(page);
2200                                 if (err == -EEXIST)
2201                                         goto retry;
2202                                 goto out;
2203                         }
2204
2205                         spin_lock(&inode->i_lock);
2206                         inode->i_blocks += blocks_per_huge_page(h);
2207                         spin_unlock(&inode->i_lock);
2208                 } else
2209                         lock_page(page);
2210         }
2211
2212         /*
2213          * If we are going to COW a private mapping later, we examine the
2214          * pending reservations for this page now. This will ensure that
2215          * any allocations necessary to record that reservation occur outside
2216          * the spinlock.
2217          */
2218         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2219                 if (vma_needs_reservation(h, vma, address) < 0) {
2220                         ret = VM_FAULT_OOM;
2221                         goto backout_unlocked;
2222                 }
2223
2224         spin_lock(&mm->page_table_lock);
2225         size = i_size_read(mapping->host) >> huge_page_shift(h);
2226         if (idx >= size)
2227                 goto backout;
2228
2229         ret = 0;
2230         if (!huge_pte_none(huge_ptep_get(ptep)))
2231                 goto backout;
2232
2233         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2234                                 && (vma->vm_flags & VM_SHARED)));
2235         set_huge_pte_at(mm, address, ptep, new_pte);
2236
2237         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2238                 /* Optimization, do the COW without a second fault */
2239                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2240         }
2241
2242         spin_unlock(&mm->page_table_lock);
2243         unlock_page(page);
2244 out:
2245         return ret;
2246
2247 backout:
2248         spin_unlock(&mm->page_table_lock);
2249 backout_unlocked:
2250         unlock_page(page);
2251         put_page(page);
2252         goto out;
2253 }
2254
2255 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2256                         unsigned long address, unsigned int flags)
2257 {
2258         pte_t *ptep;
2259         pte_t entry;
2260         int ret;
2261         struct page *pagecache_page = NULL;
2262         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2263         struct hstate *h = hstate_vma(vma);
2264
2265         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2266         if (!ptep)
2267                 return VM_FAULT_OOM;
2268
2269         /*
2270          * Serialize hugepage allocation and instantiation, so that we don't
2271          * get spurious allocation failures if two CPUs race to instantiate
2272          * the same page in the page cache.
2273          */
2274         mutex_lock(&hugetlb_instantiation_mutex);
2275         entry = huge_ptep_get(ptep);
2276         if (huge_pte_none(entry)) {
2277                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2278                 goto out_mutex;
2279         }
2280
2281         ret = 0;
2282
2283         /*
2284          * If we are going to COW the mapping later, we examine the pending
2285          * reservations for this page now. This will ensure that any
2286          * allocations necessary to record that reservation occur outside the
2287          * spinlock. For private mappings, we also lookup the pagecache
2288          * page now as it is used to determine if a reservation has been
2289          * consumed.
2290          */
2291         if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2292                 if (vma_needs_reservation(h, vma, address) < 0) {
2293                         ret = VM_FAULT_OOM;
2294                         goto out_mutex;
2295                 }
2296
2297                 if (!(vma->vm_flags & VM_MAYSHARE))
2298                         pagecache_page = hugetlbfs_pagecache_page(h,
2299                                                                 vma, address);
2300         }
2301
2302         spin_lock(&mm->page_table_lock);
2303         /* Check for a racing update before calling hugetlb_cow */
2304         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2305                 goto out_page_table_lock;
2306
2307
2308         if (flags & FAULT_FLAG_WRITE) {
2309                 if (!pte_write(entry)) {
2310                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2311                                                         pagecache_page);
2312                         goto out_page_table_lock;
2313                 }
2314                 entry = pte_mkdirty(entry);
2315         }
2316         entry = pte_mkyoung(entry);
2317         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2318                                                 flags & FAULT_FLAG_WRITE))
2319                 update_mmu_cache(vma, address, entry);
2320
2321 out_page_table_lock:
2322         spin_unlock(&mm->page_table_lock);
2323
2324         if (pagecache_page) {
2325                 unlock_page(pagecache_page);
2326                 put_page(pagecache_page);
2327         }
2328
2329 out_mutex:
2330         mutex_unlock(&hugetlb_instantiation_mutex);
2331
2332         return ret;
2333 }
2334
2335 /* Can be overriden by architectures */
2336 __attribute__((weak)) struct page *
2337 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2338                pud_t *pud, int write)
2339 {
2340         BUG();
2341         return NULL;
2342 }
2343
2344 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2345                         struct page **pages, struct vm_area_struct **vmas,
2346                         unsigned long *position, int *length, int i,
2347                         unsigned int flags)
2348 {
2349         unsigned long pfn_offset;
2350         unsigned long vaddr = *position;
2351         int remainder = *length;
2352         struct hstate *h = hstate_vma(vma);
2353
2354         spin_lock(&mm->page_table_lock);
2355         while (vaddr < vma->vm_end && remainder) {
2356                 pte_t *pte;
2357                 int absent;
2358                 struct page *page;
2359
2360                 /*
2361                  * Some archs (sparc64, sh*) have multiple pte_ts to
2362                  * each hugepage.  We have to make sure we get the
2363                  * first, for the page indexing below to work.
2364                  */
2365                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2366                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2367
2368                 /*
2369                  * When coredumping, it suits get_dump_page if we just return
2370                  * an error where there's an empty slot with no huge pagecache
2371                  * to back it.  This way, we avoid allocating a hugepage, and
2372                  * the sparse dumpfile avoids allocating disk blocks, but its
2373                  * huge holes still show up with zeroes where they need to be.
2374                  */
2375                 if (absent && (flags & FOLL_DUMP) &&
2376                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2377                         remainder = 0;
2378                         break;
2379                 }
2380
2381                 if (absent ||
2382                     ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2383                         int ret;
2384
2385                         spin_unlock(&mm->page_table_lock);
2386                         ret = hugetlb_fault(mm, vma, vaddr,
2387                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2388                         spin_lock(&mm->page_table_lock);
2389                         if (!(ret & VM_FAULT_ERROR))
2390                                 continue;
2391
2392                         remainder = 0;
2393                         break;
2394                 }
2395
2396                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2397                 page = pte_page(huge_ptep_get(pte));
2398 same_page:
2399                 if (pages) {
2400                         pages[i] = mem_map_offset(page, pfn_offset);
2401                         get_page(pages[i]);
2402                 }
2403
2404                 if (vmas)
2405                         vmas[i] = vma;
2406
2407                 vaddr += PAGE_SIZE;
2408                 ++pfn_offset;
2409                 --remainder;
2410                 ++i;
2411                 if (vaddr < vma->vm_end && remainder &&
2412                                 pfn_offset < pages_per_huge_page(h)) {
2413                         /*
2414                          * We use pfn_offset to avoid touching the pageframes
2415                          * of this compound page.
2416                          */
2417                         goto same_page;
2418                 }
2419         }
2420         spin_unlock(&mm->page_table_lock);
2421         *length = remainder;
2422         *position = vaddr;
2423
2424         return i ? i : -EFAULT;
2425 }
2426
2427 void hugetlb_change_protection(struct vm_area_struct *vma,
2428                 unsigned long address, unsigned long end, pgprot_t newprot)
2429 {
2430         struct mm_struct *mm = vma->vm_mm;
2431         unsigned long start = address;
2432         pte_t *ptep;
2433         pte_t pte;
2434         struct hstate *h = hstate_vma(vma);
2435
2436         BUG_ON(address >= end);
2437         flush_cache_range(vma, address, end);
2438
2439         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2440         spin_lock(&mm->page_table_lock);
2441         for (; address < end; address += huge_page_size(h)) {
2442                 ptep = huge_pte_offset(mm, address);
2443                 if (!ptep)
2444                         continue;
2445                 if (huge_pmd_unshare(mm, &address, ptep))
2446                         continue;
2447                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2448                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2449                         pte = pte_mkhuge(pte_modify(pte, newprot));
2450                         set_huge_pte_at(mm, address, ptep, pte);
2451                 }
2452         }
2453         spin_unlock(&mm->page_table_lock);
2454         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2455
2456         flush_tlb_range(vma, start, end);
2457 }
2458
2459 int hugetlb_reserve_pages(struct inode *inode,
2460                                         long from, long to,
2461                                         struct vm_area_struct *vma,
2462                                         int acctflag)
2463 {
2464         long ret, chg;
2465         struct hstate *h = hstate_inode(inode);
2466
2467         /*
2468          * Only apply hugepage reservation if asked. At fault time, an
2469          * attempt will be made for VM_NORESERVE to allocate a page
2470          * and filesystem quota without using reserves
2471          */
2472         if (acctflag & VM_NORESERVE)
2473                 return 0;
2474
2475         /*
2476          * Shared mappings base their reservation on the number of pages that
2477          * are already allocated on behalf of the file. Private mappings need
2478          * to reserve the full area even if read-only as mprotect() may be
2479          * called to make the mapping read-write. Assume !vma is a shm mapping
2480          */
2481         if (!vma || vma->vm_flags & VM_MAYSHARE)
2482                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2483         else {
2484                 struct resv_map *resv_map = resv_map_alloc();
2485                 if (!resv_map)
2486                         return -ENOMEM;
2487
2488                 chg = to - from;
2489
2490                 set_vma_resv_map(vma, resv_map);
2491                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2492         }
2493
2494         if (chg < 0)
2495                 return chg;
2496
2497         /* There must be enough filesystem quota for the mapping */
2498         if (hugetlb_get_quota(inode->i_mapping, chg))
2499                 return -ENOSPC;
2500
2501         /*
2502          * Check enough hugepages are available for the reservation.
2503          * Hand back the quota if there are not
2504          */
2505         ret = hugetlb_acct_memory(h, chg);
2506         if (ret < 0) {
2507                 hugetlb_put_quota(inode->i_mapping, chg);
2508                 return ret;
2509         }
2510
2511         /*
2512          * Account for the reservations made. Shared mappings record regions
2513          * that have reservations as they are shared by multiple VMAs.
2514          * When the last VMA disappears, the region map says how much
2515          * the reservation was and the page cache tells how much of
2516          * the reservation was consumed. Private mappings are per-VMA and
2517          * only the consumed reservations are tracked. When the VMA
2518          * disappears, the original reservation is the VMA size and the
2519          * consumed reservations are stored in the map. Hence, nothing
2520          * else has to be done for private mappings here
2521          */
2522         if (!vma || vma->vm_flags & VM_MAYSHARE)
2523                 region_add(&inode->i_mapping->private_list, from, to);
2524         return 0;
2525 }
2526
2527 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2528 {
2529         struct hstate *h = hstate_inode(inode);
2530         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2531
2532         spin_lock(&inode->i_lock);
2533         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2534         spin_unlock(&inode->i_lock);
2535
2536         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2537         hugetlb_acct_memory(h, -(chg - freed));
2538 }