mm: vmscan: fix force-scanning small targets without swap
[linux-2.6.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) William Irwin, April 2004
4  */
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24
25 #include <asm/page.h>
26 #include <asm/pgtable.h>
27 #include <linux/io.h>
28
29 #include <linux/hugetlb.h>
30 #include <linux/node.h>
31 #include "internal.h"
32
33 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
36
37 static int max_hstate;
38 unsigned int default_hstate_idx;
39 struct hstate hstates[HUGE_MAX_HSTATE];
40
41 __initdata LIST_HEAD(huge_boot_pages);
42
43 /* for command line parsing */
44 static struct hstate * __initdata parsed_hstate;
45 static unsigned long __initdata default_hstate_max_huge_pages;
46 static unsigned long __initdata default_hstate_size;
47
48 #define for_each_hstate(h) \
49         for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
50
51 /*
52  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53  */
54 static DEFINE_SPINLOCK(hugetlb_lock);
55
56 /*
57  * Region tracking -- allows tracking of reservations and instantiated pages
58  *                    across the pages in a mapping.
59  *
60  * The region data structures are protected by a combination of the mmap_sem
61  * and the hugetlb_instantion_mutex.  To access or modify a region the caller
62  * must either hold the mmap_sem for write, or the mmap_sem for read and
63  * the hugetlb_instantiation mutex:
64  *
65  *      down_write(&mm->mmap_sem);
66  * or
67  *      down_read(&mm->mmap_sem);
68  *      mutex_lock(&hugetlb_instantiation_mutex);
69  */
70 struct file_region {
71         struct list_head link;
72         long from;
73         long to;
74 };
75
76 static long region_add(struct list_head *head, long f, long t)
77 {
78         struct file_region *rg, *nrg, *trg;
79
80         /* Locate the region we are either in or before. */
81         list_for_each_entry(rg, head, link)
82                 if (f <= rg->to)
83                         break;
84
85         /* Round our left edge to the current segment if it encloses us. */
86         if (f > rg->from)
87                 f = rg->from;
88
89         /* Check for and consume any regions we now overlap with. */
90         nrg = rg;
91         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
92                 if (&rg->link == head)
93                         break;
94                 if (rg->from > t)
95                         break;
96
97                 /* If this area reaches higher then extend our area to
98                  * include it completely.  If this is not the first area
99                  * which we intend to reuse, free it. */
100                 if (rg->to > t)
101                         t = rg->to;
102                 if (rg != nrg) {
103                         list_del(&rg->link);
104                         kfree(rg);
105                 }
106         }
107         nrg->from = f;
108         nrg->to = t;
109         return 0;
110 }
111
112 static long region_chg(struct list_head *head, long f, long t)
113 {
114         struct file_region *rg, *nrg;
115         long chg = 0;
116
117         /* Locate the region we are before or in. */
118         list_for_each_entry(rg, head, link)
119                 if (f <= rg->to)
120                         break;
121
122         /* If we are below the current region then a new region is required.
123          * Subtle, allocate a new region at the position but make it zero
124          * size such that we can guarantee to record the reservation. */
125         if (&rg->link == head || t < rg->from) {
126                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
127                 if (!nrg)
128                         return -ENOMEM;
129                 nrg->from = f;
130                 nrg->to   = f;
131                 INIT_LIST_HEAD(&nrg->link);
132                 list_add(&nrg->link, rg->link.prev);
133
134                 return t - f;
135         }
136
137         /* Round our left edge to the current segment if it encloses us. */
138         if (f > rg->from)
139                 f = rg->from;
140         chg = t - f;
141
142         /* Check for and consume any regions we now overlap with. */
143         list_for_each_entry(rg, rg->link.prev, link) {
144                 if (&rg->link == head)
145                         break;
146                 if (rg->from > t)
147                         return chg;
148
149                 /* We overlap with this area, if it extends further than
150                  * us then we must extend ourselves.  Account for its
151                  * existing reservation. */
152                 if (rg->to > t) {
153                         chg += rg->to - t;
154                         t = rg->to;
155                 }
156                 chg -= rg->to - rg->from;
157         }
158         return chg;
159 }
160
161 static long region_truncate(struct list_head *head, long end)
162 {
163         struct file_region *rg, *trg;
164         long chg = 0;
165
166         /* Locate the region we are either in or before. */
167         list_for_each_entry(rg, head, link)
168                 if (end <= rg->to)
169                         break;
170         if (&rg->link == head)
171                 return 0;
172
173         /* If we are in the middle of a region then adjust it. */
174         if (end > rg->from) {
175                 chg = rg->to - end;
176                 rg->to = end;
177                 rg = list_entry(rg->link.next, typeof(*rg), link);
178         }
179
180         /* Drop any remaining regions. */
181         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
182                 if (&rg->link == head)
183                         break;
184                 chg += rg->to - rg->from;
185                 list_del(&rg->link);
186                 kfree(rg);
187         }
188         return chg;
189 }
190
191 static long region_count(struct list_head *head, long f, long t)
192 {
193         struct file_region *rg;
194         long chg = 0;
195
196         /* Locate each segment we overlap with, and count that overlap. */
197         list_for_each_entry(rg, head, link) {
198                 int seg_from;
199                 int seg_to;
200
201                 if (rg->to <= f)
202                         continue;
203                 if (rg->from >= t)
204                         break;
205
206                 seg_from = max(rg->from, f);
207                 seg_to = min(rg->to, t);
208
209                 chg += seg_to - seg_from;
210         }
211
212         return chg;
213 }
214
215 /*
216  * Convert the address within this vma to the page offset within
217  * the mapping, in pagecache page units; huge pages here.
218  */
219 static pgoff_t vma_hugecache_offset(struct hstate *h,
220                         struct vm_area_struct *vma, unsigned long address)
221 {
222         return ((address - vma->vm_start) >> huge_page_shift(h)) +
223                         (vma->vm_pgoff >> huge_page_order(h));
224 }
225
226 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
227                                      unsigned long address)
228 {
229         return vma_hugecache_offset(hstate_vma(vma), vma, address);
230 }
231
232 /*
233  * Return the size of the pages allocated when backing a VMA. In the majority
234  * cases this will be same size as used by the page table entries.
235  */
236 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
237 {
238         struct hstate *hstate;
239
240         if (!is_vm_hugetlb_page(vma))
241                 return PAGE_SIZE;
242
243         hstate = hstate_vma(vma);
244
245         return 1UL << (hstate->order + PAGE_SHIFT);
246 }
247 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
248
249 /*
250  * Return the page size being used by the MMU to back a VMA. In the majority
251  * of cases, the page size used by the kernel matches the MMU size. On
252  * architectures where it differs, an architecture-specific version of this
253  * function is required.
254  */
255 #ifndef vma_mmu_pagesize
256 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
257 {
258         return vma_kernel_pagesize(vma);
259 }
260 #endif
261
262 /*
263  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
264  * bits of the reservation map pointer, which are always clear due to
265  * alignment.
266  */
267 #define HPAGE_RESV_OWNER    (1UL << 0)
268 #define HPAGE_RESV_UNMAPPED (1UL << 1)
269 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
270
271 /*
272  * These helpers are used to track how many pages are reserved for
273  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
274  * is guaranteed to have their future faults succeed.
275  *
276  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
277  * the reserve counters are updated with the hugetlb_lock held. It is safe
278  * to reset the VMA at fork() time as it is not in use yet and there is no
279  * chance of the global counters getting corrupted as a result of the values.
280  *
281  * The private mapping reservation is represented in a subtly different
282  * manner to a shared mapping.  A shared mapping has a region map associated
283  * with the underlying file, this region map represents the backing file
284  * pages which have ever had a reservation assigned which this persists even
285  * after the page is instantiated.  A private mapping has a region map
286  * associated with the original mmap which is attached to all VMAs which
287  * reference it, this region map represents those offsets which have consumed
288  * reservation ie. where pages have been instantiated.
289  */
290 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
291 {
292         return (unsigned long)vma->vm_private_data;
293 }
294
295 static void set_vma_private_data(struct vm_area_struct *vma,
296                                                         unsigned long value)
297 {
298         vma->vm_private_data = (void *)value;
299 }
300
301 struct resv_map {
302         struct kref refs;
303         struct list_head regions;
304 };
305
306 static struct resv_map *resv_map_alloc(void)
307 {
308         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
309         if (!resv_map)
310                 return NULL;
311
312         kref_init(&resv_map->refs);
313         INIT_LIST_HEAD(&resv_map->regions);
314
315         return resv_map;
316 }
317
318 static void resv_map_release(struct kref *ref)
319 {
320         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
321
322         /* Clear out any active regions before we release the map. */
323         region_truncate(&resv_map->regions, 0);
324         kfree(resv_map);
325 }
326
327 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
328 {
329         VM_BUG_ON(!is_vm_hugetlb_page(vma));
330         if (!(vma->vm_flags & VM_MAYSHARE))
331                 return (struct resv_map *)(get_vma_private_data(vma) &
332                                                         ~HPAGE_RESV_MASK);
333         return NULL;
334 }
335
336 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
337 {
338         VM_BUG_ON(!is_vm_hugetlb_page(vma));
339         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
340
341         set_vma_private_data(vma, (get_vma_private_data(vma) &
342                                 HPAGE_RESV_MASK) | (unsigned long)map);
343 }
344
345 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
346 {
347         VM_BUG_ON(!is_vm_hugetlb_page(vma));
348         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
349
350         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
351 }
352
353 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
354 {
355         VM_BUG_ON(!is_vm_hugetlb_page(vma));
356
357         return (get_vma_private_data(vma) & flag) != 0;
358 }
359
360 /* Decrement the reserved pages in the hugepage pool by one */
361 static void decrement_hugepage_resv_vma(struct hstate *h,
362                         struct vm_area_struct *vma)
363 {
364         if (vma->vm_flags & VM_NORESERVE)
365                 return;
366
367         if (vma->vm_flags & VM_MAYSHARE) {
368                 /* Shared mappings always use reserves */
369                 h->resv_huge_pages--;
370         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
371                 /*
372                  * Only the process that called mmap() has reserves for
373                  * private mappings.
374                  */
375                 h->resv_huge_pages--;
376         }
377 }
378
379 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
380 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
381 {
382         VM_BUG_ON(!is_vm_hugetlb_page(vma));
383         if (!(vma->vm_flags & VM_MAYSHARE))
384                 vma->vm_private_data = (void *)0;
385 }
386
387 /* Returns true if the VMA has associated reserve pages */
388 static int vma_has_reserves(struct vm_area_struct *vma)
389 {
390         if (vma->vm_flags & VM_MAYSHARE)
391                 return 1;
392         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
393                 return 1;
394         return 0;
395 }
396
397 static void copy_gigantic_page(struct page *dst, struct page *src)
398 {
399         int i;
400         struct hstate *h = page_hstate(src);
401         struct page *dst_base = dst;
402         struct page *src_base = src;
403
404         for (i = 0; i < pages_per_huge_page(h); ) {
405                 cond_resched();
406                 copy_highpage(dst, src);
407
408                 i++;
409                 dst = mem_map_next(dst, dst_base, i);
410                 src = mem_map_next(src, src_base, i);
411         }
412 }
413
414 void copy_huge_page(struct page *dst, struct page *src)
415 {
416         int i;
417         struct hstate *h = page_hstate(src);
418
419         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
420                 copy_gigantic_page(dst, src);
421                 return;
422         }
423
424         might_sleep();
425         for (i = 0; i < pages_per_huge_page(h); i++) {
426                 cond_resched();
427                 copy_highpage(dst + i, src + i);
428         }
429 }
430
431 static void enqueue_huge_page(struct hstate *h, struct page *page)
432 {
433         int nid = page_to_nid(page);
434         list_add(&page->lru, &h->hugepage_freelists[nid]);
435         h->free_huge_pages++;
436         h->free_huge_pages_node[nid]++;
437 }
438
439 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
440 {
441         struct page *page;
442
443         if (list_empty(&h->hugepage_freelists[nid]))
444                 return NULL;
445         page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
446         list_del(&page->lru);
447         set_page_refcounted(page);
448         h->free_huge_pages--;
449         h->free_huge_pages_node[nid]--;
450         return page;
451 }
452
453 static struct page *dequeue_huge_page_vma(struct hstate *h,
454                                 struct vm_area_struct *vma,
455                                 unsigned long address, int avoid_reserve)
456 {
457         struct page *page = NULL;
458         struct mempolicy *mpol;
459         nodemask_t *nodemask;
460         struct zonelist *zonelist;
461         struct zone *zone;
462         struct zoneref *z;
463
464         get_mems_allowed();
465         zonelist = huge_zonelist(vma, address,
466                                         htlb_alloc_mask, &mpol, &nodemask);
467         /*
468          * A child process with MAP_PRIVATE mappings created by their parent
469          * have no page reserves. This check ensures that reservations are
470          * not "stolen". The child may still get SIGKILLed
471          */
472         if (!vma_has_reserves(vma) &&
473                         h->free_huge_pages - h->resv_huge_pages == 0)
474                 goto err;
475
476         /* If reserves cannot be used, ensure enough pages are in the pool */
477         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
478                 goto err;
479
480         for_each_zone_zonelist_nodemask(zone, z, zonelist,
481                                                 MAX_NR_ZONES - 1, nodemask) {
482                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
483                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
484                         if (page) {
485                                 if (!avoid_reserve)
486                                         decrement_hugepage_resv_vma(h, vma);
487                                 break;
488                         }
489                 }
490         }
491 err:
492         mpol_cond_put(mpol);
493         put_mems_allowed();
494         return page;
495 }
496
497 static void update_and_free_page(struct hstate *h, struct page *page)
498 {
499         int i;
500
501         VM_BUG_ON(h->order >= MAX_ORDER);
502
503         h->nr_huge_pages--;
504         h->nr_huge_pages_node[page_to_nid(page)]--;
505         for (i = 0; i < pages_per_huge_page(h); i++) {
506                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
507                                 1 << PG_referenced | 1 << PG_dirty |
508                                 1 << PG_active | 1 << PG_reserved |
509                                 1 << PG_private | 1 << PG_writeback);
510         }
511         set_compound_page_dtor(page, NULL);
512         set_page_refcounted(page);
513         arch_release_hugepage(page);
514         __free_pages(page, huge_page_order(h));
515 }
516
517 struct hstate *size_to_hstate(unsigned long size)
518 {
519         struct hstate *h;
520
521         for_each_hstate(h) {
522                 if (huge_page_size(h) == size)
523                         return h;
524         }
525         return NULL;
526 }
527
528 static void free_huge_page(struct page *page)
529 {
530         /*
531          * Can't pass hstate in here because it is called from the
532          * compound page destructor.
533          */
534         struct hstate *h = page_hstate(page);
535         int nid = page_to_nid(page);
536         struct address_space *mapping;
537
538         mapping = (struct address_space *) page_private(page);
539         set_page_private(page, 0);
540         page->mapping = NULL;
541         BUG_ON(page_count(page));
542         BUG_ON(page_mapcount(page));
543         INIT_LIST_HEAD(&page->lru);
544
545         spin_lock(&hugetlb_lock);
546         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
547                 update_and_free_page(h, page);
548                 h->surplus_huge_pages--;
549                 h->surplus_huge_pages_node[nid]--;
550         } else {
551                 enqueue_huge_page(h, page);
552         }
553         spin_unlock(&hugetlb_lock);
554         if (mapping)
555                 hugetlb_put_quota(mapping, 1);
556 }
557
558 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
559 {
560         set_compound_page_dtor(page, free_huge_page);
561         spin_lock(&hugetlb_lock);
562         h->nr_huge_pages++;
563         h->nr_huge_pages_node[nid]++;
564         spin_unlock(&hugetlb_lock);
565         put_page(page); /* free it into the hugepage allocator */
566 }
567
568 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
569 {
570         int i;
571         int nr_pages = 1 << order;
572         struct page *p = page + 1;
573
574         /* we rely on prep_new_huge_page to set the destructor */
575         set_compound_order(page, order);
576         __SetPageHead(page);
577         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
578                 __SetPageTail(p);
579                 p->first_page = page;
580         }
581 }
582
583 int PageHuge(struct page *page)
584 {
585         compound_page_dtor *dtor;
586
587         if (!PageCompound(page))
588                 return 0;
589
590         page = compound_head(page);
591         dtor = get_compound_page_dtor(page);
592
593         return dtor == free_huge_page;
594 }
595 EXPORT_SYMBOL_GPL(PageHuge);
596
597 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
598 {
599         struct page *page;
600
601         if (h->order >= MAX_ORDER)
602                 return NULL;
603
604         page = alloc_pages_exact_node(nid,
605                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
606                                                 __GFP_REPEAT|__GFP_NOWARN,
607                 huge_page_order(h));
608         if (page) {
609                 if (arch_prepare_hugepage(page)) {
610                         __free_pages(page, huge_page_order(h));
611                         return NULL;
612                 }
613                 prep_new_huge_page(h, page, nid);
614         }
615
616         return page;
617 }
618
619 /*
620  * common helper functions for hstate_next_node_to_{alloc|free}.
621  * We may have allocated or freed a huge page based on a different
622  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
623  * be outside of *nodes_allowed.  Ensure that we use an allowed
624  * node for alloc or free.
625  */
626 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
627 {
628         nid = next_node(nid, *nodes_allowed);
629         if (nid == MAX_NUMNODES)
630                 nid = first_node(*nodes_allowed);
631         VM_BUG_ON(nid >= MAX_NUMNODES);
632
633         return nid;
634 }
635
636 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
637 {
638         if (!node_isset(nid, *nodes_allowed))
639                 nid = next_node_allowed(nid, nodes_allowed);
640         return nid;
641 }
642
643 /*
644  * returns the previously saved node ["this node"] from which to
645  * allocate a persistent huge page for the pool and advance the
646  * next node from which to allocate, handling wrap at end of node
647  * mask.
648  */
649 static int hstate_next_node_to_alloc(struct hstate *h,
650                                         nodemask_t *nodes_allowed)
651 {
652         int nid;
653
654         VM_BUG_ON(!nodes_allowed);
655
656         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
657         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
658
659         return nid;
660 }
661
662 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
663 {
664         struct page *page;
665         int start_nid;
666         int next_nid;
667         int ret = 0;
668
669         start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
670         next_nid = start_nid;
671
672         do {
673                 page = alloc_fresh_huge_page_node(h, next_nid);
674                 if (page) {
675                         ret = 1;
676                         break;
677                 }
678                 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
679         } while (next_nid != start_nid);
680
681         if (ret)
682                 count_vm_event(HTLB_BUDDY_PGALLOC);
683         else
684                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
685
686         return ret;
687 }
688
689 /*
690  * helper for free_pool_huge_page() - return the previously saved
691  * node ["this node"] from which to free a huge page.  Advance the
692  * next node id whether or not we find a free huge page to free so
693  * that the next attempt to free addresses the next node.
694  */
695 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
696 {
697         int nid;
698
699         VM_BUG_ON(!nodes_allowed);
700
701         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
702         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
703
704         return nid;
705 }
706
707 /*
708  * Free huge page from pool from next node to free.
709  * Attempt to keep persistent huge pages more or less
710  * balanced over allowed nodes.
711  * Called with hugetlb_lock locked.
712  */
713 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
714                                                          bool acct_surplus)
715 {
716         int start_nid;
717         int next_nid;
718         int ret = 0;
719
720         start_nid = hstate_next_node_to_free(h, nodes_allowed);
721         next_nid = start_nid;
722
723         do {
724                 /*
725                  * If we're returning unused surplus pages, only examine
726                  * nodes with surplus pages.
727                  */
728                 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
729                     !list_empty(&h->hugepage_freelists[next_nid])) {
730                         struct page *page =
731                                 list_entry(h->hugepage_freelists[next_nid].next,
732                                           struct page, lru);
733                         list_del(&page->lru);
734                         h->free_huge_pages--;
735                         h->free_huge_pages_node[next_nid]--;
736                         if (acct_surplus) {
737                                 h->surplus_huge_pages--;
738                                 h->surplus_huge_pages_node[next_nid]--;
739                         }
740                         update_and_free_page(h, page);
741                         ret = 1;
742                         break;
743                 }
744                 next_nid = hstate_next_node_to_free(h, nodes_allowed);
745         } while (next_nid != start_nid);
746
747         return ret;
748 }
749
750 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
751 {
752         struct page *page;
753         unsigned int r_nid;
754
755         if (h->order >= MAX_ORDER)
756                 return NULL;
757
758         /*
759          * Assume we will successfully allocate the surplus page to
760          * prevent racing processes from causing the surplus to exceed
761          * overcommit
762          *
763          * This however introduces a different race, where a process B
764          * tries to grow the static hugepage pool while alloc_pages() is
765          * called by process A. B will only examine the per-node
766          * counters in determining if surplus huge pages can be
767          * converted to normal huge pages in adjust_pool_surplus(). A
768          * won't be able to increment the per-node counter, until the
769          * lock is dropped by B, but B doesn't drop hugetlb_lock until
770          * no more huge pages can be converted from surplus to normal
771          * state (and doesn't try to convert again). Thus, we have a
772          * case where a surplus huge page exists, the pool is grown, and
773          * the surplus huge page still exists after, even though it
774          * should just have been converted to a normal huge page. This
775          * does not leak memory, though, as the hugepage will be freed
776          * once it is out of use. It also does not allow the counters to
777          * go out of whack in adjust_pool_surplus() as we don't modify
778          * the node values until we've gotten the hugepage and only the
779          * per-node value is checked there.
780          */
781         spin_lock(&hugetlb_lock);
782         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
783                 spin_unlock(&hugetlb_lock);
784                 return NULL;
785         } else {
786                 h->nr_huge_pages++;
787                 h->surplus_huge_pages++;
788         }
789         spin_unlock(&hugetlb_lock);
790
791         if (nid == NUMA_NO_NODE)
792                 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
793                                    __GFP_REPEAT|__GFP_NOWARN,
794                                    huge_page_order(h));
795         else
796                 page = alloc_pages_exact_node(nid,
797                         htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
798                         __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
799
800         if (page && arch_prepare_hugepage(page)) {
801                 __free_pages(page, huge_page_order(h));
802                 return NULL;
803         }
804
805         spin_lock(&hugetlb_lock);
806         if (page) {
807                 r_nid = page_to_nid(page);
808                 set_compound_page_dtor(page, free_huge_page);
809                 /*
810                  * We incremented the global counters already
811                  */
812                 h->nr_huge_pages_node[r_nid]++;
813                 h->surplus_huge_pages_node[r_nid]++;
814                 __count_vm_event(HTLB_BUDDY_PGALLOC);
815         } else {
816                 h->nr_huge_pages--;
817                 h->surplus_huge_pages--;
818                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
819         }
820         spin_unlock(&hugetlb_lock);
821
822         return page;
823 }
824
825 /*
826  * This allocation function is useful in the context where vma is irrelevant.
827  * E.g. soft-offlining uses this function because it only cares physical
828  * address of error page.
829  */
830 struct page *alloc_huge_page_node(struct hstate *h, int nid)
831 {
832         struct page *page;
833
834         spin_lock(&hugetlb_lock);
835         page = dequeue_huge_page_node(h, nid);
836         spin_unlock(&hugetlb_lock);
837
838         if (!page)
839                 page = alloc_buddy_huge_page(h, nid);
840
841         return page;
842 }
843
844 /*
845  * Increase the hugetlb pool such that it can accommodate a reservation
846  * of size 'delta'.
847  */
848 static int gather_surplus_pages(struct hstate *h, int delta)
849 {
850         struct list_head surplus_list;
851         struct page *page, *tmp;
852         int ret, i;
853         int needed, allocated;
854
855         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
856         if (needed <= 0) {
857                 h->resv_huge_pages += delta;
858                 return 0;
859         }
860
861         allocated = 0;
862         INIT_LIST_HEAD(&surplus_list);
863
864         ret = -ENOMEM;
865 retry:
866         spin_unlock(&hugetlb_lock);
867         for (i = 0; i < needed; i++) {
868                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
869                 if (!page)
870                         /*
871                          * We were not able to allocate enough pages to
872                          * satisfy the entire reservation so we free what
873                          * we've allocated so far.
874                          */
875                         goto free;
876
877                 list_add(&page->lru, &surplus_list);
878         }
879         allocated += needed;
880
881         /*
882          * After retaking hugetlb_lock, we need to recalculate 'needed'
883          * because either resv_huge_pages or free_huge_pages may have changed.
884          */
885         spin_lock(&hugetlb_lock);
886         needed = (h->resv_huge_pages + delta) -
887                         (h->free_huge_pages + allocated);
888         if (needed > 0)
889                 goto retry;
890
891         /*
892          * The surplus_list now contains _at_least_ the number of extra pages
893          * needed to accommodate the reservation.  Add the appropriate number
894          * of pages to the hugetlb pool and free the extras back to the buddy
895          * allocator.  Commit the entire reservation here to prevent another
896          * process from stealing the pages as they are added to the pool but
897          * before they are reserved.
898          */
899         needed += allocated;
900         h->resv_huge_pages += delta;
901         ret = 0;
902
903         spin_unlock(&hugetlb_lock);
904         /* Free the needed pages to the hugetlb pool */
905         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
906                 if ((--needed) < 0)
907                         break;
908                 list_del(&page->lru);
909                 /*
910                  * This page is now managed by the hugetlb allocator and has
911                  * no users -- drop the buddy allocator's reference.
912                  */
913                 put_page_testzero(page);
914                 VM_BUG_ON(page_count(page));
915                 enqueue_huge_page(h, page);
916         }
917
918         /* Free unnecessary surplus pages to the buddy allocator */
919 free:
920         if (!list_empty(&surplus_list)) {
921                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
922                         list_del(&page->lru);
923                         put_page(page);
924                 }
925         }
926         spin_lock(&hugetlb_lock);
927
928         return ret;
929 }
930
931 /*
932  * When releasing a hugetlb pool reservation, any surplus pages that were
933  * allocated to satisfy the reservation must be explicitly freed if they were
934  * never used.
935  * Called with hugetlb_lock held.
936  */
937 static void return_unused_surplus_pages(struct hstate *h,
938                                         unsigned long unused_resv_pages)
939 {
940         unsigned long nr_pages;
941
942         /* Uncommit the reservation */
943         h->resv_huge_pages -= unused_resv_pages;
944
945         /* Cannot return gigantic pages currently */
946         if (h->order >= MAX_ORDER)
947                 return;
948
949         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
950
951         /*
952          * We want to release as many surplus pages as possible, spread
953          * evenly across all nodes with memory. Iterate across these nodes
954          * until we can no longer free unreserved surplus pages. This occurs
955          * when the nodes with surplus pages have no free pages.
956          * free_pool_huge_page() will balance the the freed pages across the
957          * on-line nodes with memory and will handle the hstate accounting.
958          */
959         while (nr_pages--) {
960                 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
961                         break;
962         }
963 }
964
965 /*
966  * Determine if the huge page at addr within the vma has an associated
967  * reservation.  Where it does not we will need to logically increase
968  * reservation and actually increase quota before an allocation can occur.
969  * Where any new reservation would be required the reservation change is
970  * prepared, but not committed.  Once the page has been quota'd allocated
971  * an instantiated the change should be committed via vma_commit_reservation.
972  * No action is required on failure.
973  */
974 static long vma_needs_reservation(struct hstate *h,
975                         struct vm_area_struct *vma, unsigned long addr)
976 {
977         struct address_space *mapping = vma->vm_file->f_mapping;
978         struct inode *inode = mapping->host;
979
980         if (vma->vm_flags & VM_MAYSHARE) {
981                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
982                 return region_chg(&inode->i_mapping->private_list,
983                                                         idx, idx + 1);
984
985         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
986                 return 1;
987
988         } else  {
989                 long err;
990                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
991                 struct resv_map *reservations = vma_resv_map(vma);
992
993                 err = region_chg(&reservations->regions, idx, idx + 1);
994                 if (err < 0)
995                         return err;
996                 return 0;
997         }
998 }
999 static void vma_commit_reservation(struct hstate *h,
1000                         struct vm_area_struct *vma, unsigned long addr)
1001 {
1002         struct address_space *mapping = vma->vm_file->f_mapping;
1003         struct inode *inode = mapping->host;
1004
1005         if (vma->vm_flags & VM_MAYSHARE) {
1006                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1007                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1008
1009         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1010                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1011                 struct resv_map *reservations = vma_resv_map(vma);
1012
1013                 /* Mark this page used in the map. */
1014                 region_add(&reservations->regions, idx, idx + 1);
1015         }
1016 }
1017
1018 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1019                                     unsigned long addr, int avoid_reserve)
1020 {
1021         struct hstate *h = hstate_vma(vma);
1022         struct page *page;
1023         struct address_space *mapping = vma->vm_file->f_mapping;
1024         struct inode *inode = mapping->host;
1025         long chg;
1026
1027         /*
1028          * Processes that did not create the mapping will have no reserves and
1029          * will not have accounted against quota. Check that the quota can be
1030          * made before satisfying the allocation
1031          * MAP_NORESERVE mappings may also need pages and quota allocated
1032          * if no reserve mapping overlaps.
1033          */
1034         chg = vma_needs_reservation(h, vma, addr);
1035         if (chg < 0)
1036                 return ERR_PTR(-VM_FAULT_OOM);
1037         if (chg)
1038                 if (hugetlb_get_quota(inode->i_mapping, chg))
1039                         return ERR_PTR(-VM_FAULT_SIGBUS);
1040
1041         spin_lock(&hugetlb_lock);
1042         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1043         spin_unlock(&hugetlb_lock);
1044
1045         if (!page) {
1046                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1047                 if (!page) {
1048                         hugetlb_put_quota(inode->i_mapping, chg);
1049                         return ERR_PTR(-VM_FAULT_SIGBUS);
1050                 }
1051         }
1052
1053         set_page_private(page, (unsigned long) mapping);
1054
1055         vma_commit_reservation(h, vma, addr);
1056
1057         return page;
1058 }
1059
1060 int __weak alloc_bootmem_huge_page(struct hstate *h)
1061 {
1062         struct huge_bootmem_page *m;
1063         int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1064
1065         while (nr_nodes) {
1066                 void *addr;
1067
1068                 addr = __alloc_bootmem_node_nopanic(
1069                                 NODE_DATA(hstate_next_node_to_alloc(h,
1070                                                 &node_states[N_HIGH_MEMORY])),
1071                                 huge_page_size(h), huge_page_size(h), 0);
1072
1073                 if (addr) {
1074                         /*
1075                          * Use the beginning of the huge page to store the
1076                          * huge_bootmem_page struct (until gather_bootmem
1077                          * puts them into the mem_map).
1078                          */
1079                         m = addr;
1080                         goto found;
1081                 }
1082                 nr_nodes--;
1083         }
1084         return 0;
1085
1086 found:
1087         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1088         /* Put them into a private list first because mem_map is not up yet */
1089         list_add(&m->list, &huge_boot_pages);
1090         m->hstate = h;
1091         return 1;
1092 }
1093
1094 static void prep_compound_huge_page(struct page *page, int order)
1095 {
1096         if (unlikely(order > (MAX_ORDER - 1)))
1097                 prep_compound_gigantic_page(page, order);
1098         else
1099                 prep_compound_page(page, order);
1100 }
1101
1102 /* Put bootmem huge pages into the standard lists after mem_map is up */
1103 static void __init gather_bootmem_prealloc(void)
1104 {
1105         struct huge_bootmem_page *m;
1106
1107         list_for_each_entry(m, &huge_boot_pages, list) {
1108                 struct hstate *h = m->hstate;
1109                 struct page *page;
1110
1111 #ifdef CONFIG_HIGHMEM
1112                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1113                 free_bootmem_late((unsigned long)m,
1114                                   sizeof(struct huge_bootmem_page));
1115 #else
1116                 page = virt_to_page(m);
1117 #endif
1118                 __ClearPageReserved(page);
1119                 WARN_ON(page_count(page) != 1);
1120                 prep_compound_huge_page(page, h->order);
1121                 prep_new_huge_page(h, page, page_to_nid(page));
1122                 /*
1123                  * If we had gigantic hugepages allocated at boot time, we need
1124                  * to restore the 'stolen' pages to totalram_pages in order to
1125                  * fix confusing memory reports from free(1) and another
1126                  * side-effects, like CommitLimit going negative.
1127                  */
1128                 if (h->order > (MAX_ORDER - 1))
1129                         totalram_pages += 1 << h->order;
1130         }
1131 }
1132
1133 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1134 {
1135         unsigned long i;
1136
1137         for (i = 0; i < h->max_huge_pages; ++i) {
1138                 if (h->order >= MAX_ORDER) {
1139                         if (!alloc_bootmem_huge_page(h))
1140                                 break;
1141                 } else if (!alloc_fresh_huge_page(h,
1142                                          &node_states[N_HIGH_MEMORY]))
1143                         break;
1144         }
1145         h->max_huge_pages = i;
1146 }
1147
1148 static void __init hugetlb_init_hstates(void)
1149 {
1150         struct hstate *h;
1151
1152         for_each_hstate(h) {
1153                 /* oversize hugepages were init'ed in early boot */
1154                 if (h->order < MAX_ORDER)
1155                         hugetlb_hstate_alloc_pages(h);
1156         }
1157 }
1158
1159 static char * __init memfmt(char *buf, unsigned long n)
1160 {
1161         if (n >= (1UL << 30))
1162                 sprintf(buf, "%lu GB", n >> 30);
1163         else if (n >= (1UL << 20))
1164                 sprintf(buf, "%lu MB", n >> 20);
1165         else
1166                 sprintf(buf, "%lu KB", n >> 10);
1167         return buf;
1168 }
1169
1170 static void __init report_hugepages(void)
1171 {
1172         struct hstate *h;
1173
1174         for_each_hstate(h) {
1175                 char buf[32];
1176                 printk(KERN_INFO "HugeTLB registered %s page size, "
1177                                  "pre-allocated %ld pages\n",
1178                         memfmt(buf, huge_page_size(h)),
1179                         h->free_huge_pages);
1180         }
1181 }
1182
1183 #ifdef CONFIG_HIGHMEM
1184 static void try_to_free_low(struct hstate *h, unsigned long count,
1185                                                 nodemask_t *nodes_allowed)
1186 {
1187         int i;
1188
1189         if (h->order >= MAX_ORDER)
1190                 return;
1191
1192         for_each_node_mask(i, *nodes_allowed) {
1193                 struct page *page, *next;
1194                 struct list_head *freel = &h->hugepage_freelists[i];
1195                 list_for_each_entry_safe(page, next, freel, lru) {
1196                         if (count >= h->nr_huge_pages)
1197                                 return;
1198                         if (PageHighMem(page))
1199                                 continue;
1200                         list_del(&page->lru);
1201                         update_and_free_page(h, page);
1202                         h->free_huge_pages--;
1203                         h->free_huge_pages_node[page_to_nid(page)]--;
1204                 }
1205         }
1206 }
1207 #else
1208 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1209                                                 nodemask_t *nodes_allowed)
1210 {
1211 }
1212 #endif
1213
1214 /*
1215  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1216  * balanced by operating on them in a round-robin fashion.
1217  * Returns 1 if an adjustment was made.
1218  */
1219 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1220                                 int delta)
1221 {
1222         int start_nid, next_nid;
1223         int ret = 0;
1224
1225         VM_BUG_ON(delta != -1 && delta != 1);
1226
1227         if (delta < 0)
1228                 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1229         else
1230                 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1231         next_nid = start_nid;
1232
1233         do {
1234                 int nid = next_nid;
1235                 if (delta < 0)  {
1236                         /*
1237                          * To shrink on this node, there must be a surplus page
1238                          */
1239                         if (!h->surplus_huge_pages_node[nid]) {
1240                                 next_nid = hstate_next_node_to_alloc(h,
1241                                                                 nodes_allowed);
1242                                 continue;
1243                         }
1244                 }
1245                 if (delta > 0) {
1246                         /*
1247                          * Surplus cannot exceed the total number of pages
1248                          */
1249                         if (h->surplus_huge_pages_node[nid] >=
1250                                                 h->nr_huge_pages_node[nid]) {
1251                                 next_nid = hstate_next_node_to_free(h,
1252                                                                 nodes_allowed);
1253                                 continue;
1254                         }
1255                 }
1256
1257                 h->surplus_huge_pages += delta;
1258                 h->surplus_huge_pages_node[nid] += delta;
1259                 ret = 1;
1260                 break;
1261         } while (next_nid != start_nid);
1262
1263         return ret;
1264 }
1265
1266 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1267 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1268                                                 nodemask_t *nodes_allowed)
1269 {
1270         unsigned long min_count, ret;
1271
1272         if (h->order >= MAX_ORDER)
1273                 return h->max_huge_pages;
1274
1275         /*
1276          * Increase the pool size
1277          * First take pages out of surplus state.  Then make up the
1278          * remaining difference by allocating fresh huge pages.
1279          *
1280          * We might race with alloc_buddy_huge_page() here and be unable
1281          * to convert a surplus huge page to a normal huge page. That is
1282          * not critical, though, it just means the overall size of the
1283          * pool might be one hugepage larger than it needs to be, but
1284          * within all the constraints specified by the sysctls.
1285          */
1286         spin_lock(&hugetlb_lock);
1287         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1288                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1289                         break;
1290         }
1291
1292         while (count > persistent_huge_pages(h)) {
1293                 /*
1294                  * If this allocation races such that we no longer need the
1295                  * page, free_huge_page will handle it by freeing the page
1296                  * and reducing the surplus.
1297                  */
1298                 spin_unlock(&hugetlb_lock);
1299                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1300                 spin_lock(&hugetlb_lock);
1301                 if (!ret)
1302                         goto out;
1303
1304                 /* Bail for signals. Probably ctrl-c from user */
1305                 if (signal_pending(current))
1306                         goto out;
1307         }
1308
1309         /*
1310          * Decrease the pool size
1311          * First return free pages to the buddy allocator (being careful
1312          * to keep enough around to satisfy reservations).  Then place
1313          * pages into surplus state as needed so the pool will shrink
1314          * to the desired size as pages become free.
1315          *
1316          * By placing pages into the surplus state independent of the
1317          * overcommit value, we are allowing the surplus pool size to
1318          * exceed overcommit. There are few sane options here. Since
1319          * alloc_buddy_huge_page() is checking the global counter,
1320          * though, we'll note that we're not allowed to exceed surplus
1321          * and won't grow the pool anywhere else. Not until one of the
1322          * sysctls are changed, or the surplus pages go out of use.
1323          */
1324         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1325         min_count = max(count, min_count);
1326         try_to_free_low(h, min_count, nodes_allowed);
1327         while (min_count < persistent_huge_pages(h)) {
1328                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1329                         break;
1330         }
1331         while (count < persistent_huge_pages(h)) {
1332                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1333                         break;
1334         }
1335 out:
1336         ret = persistent_huge_pages(h);
1337         spin_unlock(&hugetlb_lock);
1338         return ret;
1339 }
1340
1341 #define HSTATE_ATTR_RO(_name) \
1342         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1343
1344 #define HSTATE_ATTR(_name) \
1345         static struct kobj_attribute _name##_attr = \
1346                 __ATTR(_name, 0644, _name##_show, _name##_store)
1347
1348 static struct kobject *hugepages_kobj;
1349 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1350
1351 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1352
1353 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1354 {
1355         int i;
1356
1357         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1358                 if (hstate_kobjs[i] == kobj) {
1359                         if (nidp)
1360                                 *nidp = NUMA_NO_NODE;
1361                         return &hstates[i];
1362                 }
1363
1364         return kobj_to_node_hstate(kobj, nidp);
1365 }
1366
1367 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1368                                         struct kobj_attribute *attr, char *buf)
1369 {
1370         struct hstate *h;
1371         unsigned long nr_huge_pages;
1372         int nid;
1373
1374         h = kobj_to_hstate(kobj, &nid);
1375         if (nid == NUMA_NO_NODE)
1376                 nr_huge_pages = h->nr_huge_pages;
1377         else
1378                 nr_huge_pages = h->nr_huge_pages_node[nid];
1379
1380         return sprintf(buf, "%lu\n", nr_huge_pages);
1381 }
1382
1383 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1384                         struct kobject *kobj, struct kobj_attribute *attr,
1385                         const char *buf, size_t len)
1386 {
1387         int err;
1388         int nid;
1389         unsigned long count;
1390         struct hstate *h;
1391         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1392
1393         err = strict_strtoul(buf, 10, &count);
1394         if (err)
1395                 goto out;
1396
1397         h = kobj_to_hstate(kobj, &nid);
1398         if (h->order >= MAX_ORDER) {
1399                 err = -EINVAL;
1400                 goto out;
1401         }
1402
1403         if (nid == NUMA_NO_NODE) {
1404                 /*
1405                  * global hstate attribute
1406                  */
1407                 if (!(obey_mempolicy &&
1408                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1409                         NODEMASK_FREE(nodes_allowed);
1410                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1411                 }
1412         } else if (nodes_allowed) {
1413                 /*
1414                  * per node hstate attribute: adjust count to global,
1415                  * but restrict alloc/free to the specified node.
1416                  */
1417                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1418                 init_nodemask_of_node(nodes_allowed, nid);
1419         } else
1420                 nodes_allowed = &node_states[N_HIGH_MEMORY];
1421
1422         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1423
1424         if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1425                 NODEMASK_FREE(nodes_allowed);
1426
1427         return len;
1428 out:
1429         NODEMASK_FREE(nodes_allowed);
1430         return err;
1431 }
1432
1433 static ssize_t nr_hugepages_show(struct kobject *kobj,
1434                                        struct kobj_attribute *attr, char *buf)
1435 {
1436         return nr_hugepages_show_common(kobj, attr, buf);
1437 }
1438
1439 static ssize_t nr_hugepages_store(struct kobject *kobj,
1440                struct kobj_attribute *attr, const char *buf, size_t len)
1441 {
1442         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1443 }
1444 HSTATE_ATTR(nr_hugepages);
1445
1446 #ifdef CONFIG_NUMA
1447
1448 /*
1449  * hstate attribute for optionally mempolicy-based constraint on persistent
1450  * huge page alloc/free.
1451  */
1452 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1453                                        struct kobj_attribute *attr, char *buf)
1454 {
1455         return nr_hugepages_show_common(kobj, attr, buf);
1456 }
1457
1458 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1459                struct kobj_attribute *attr, const char *buf, size_t len)
1460 {
1461         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1462 }
1463 HSTATE_ATTR(nr_hugepages_mempolicy);
1464 #endif
1465
1466
1467 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1468                                         struct kobj_attribute *attr, char *buf)
1469 {
1470         struct hstate *h = kobj_to_hstate(kobj, NULL);
1471         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1472 }
1473
1474 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1475                 struct kobj_attribute *attr, const char *buf, size_t count)
1476 {
1477         int err;
1478         unsigned long input;
1479         struct hstate *h = kobj_to_hstate(kobj, NULL);
1480
1481         if (h->order >= MAX_ORDER)
1482                 return -EINVAL;
1483
1484         err = strict_strtoul(buf, 10, &input);
1485         if (err)
1486                 return err;
1487
1488         spin_lock(&hugetlb_lock);
1489         h->nr_overcommit_huge_pages = input;
1490         spin_unlock(&hugetlb_lock);
1491
1492         return count;
1493 }
1494 HSTATE_ATTR(nr_overcommit_hugepages);
1495
1496 static ssize_t free_hugepages_show(struct kobject *kobj,
1497                                         struct kobj_attribute *attr, char *buf)
1498 {
1499         struct hstate *h;
1500         unsigned long free_huge_pages;
1501         int nid;
1502
1503         h = kobj_to_hstate(kobj, &nid);
1504         if (nid == NUMA_NO_NODE)
1505                 free_huge_pages = h->free_huge_pages;
1506         else
1507                 free_huge_pages = h->free_huge_pages_node[nid];
1508
1509         return sprintf(buf, "%lu\n", free_huge_pages);
1510 }
1511 HSTATE_ATTR_RO(free_hugepages);
1512
1513 static ssize_t resv_hugepages_show(struct kobject *kobj,
1514                                         struct kobj_attribute *attr, char *buf)
1515 {
1516         struct hstate *h = kobj_to_hstate(kobj, NULL);
1517         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1518 }
1519 HSTATE_ATTR_RO(resv_hugepages);
1520
1521 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1522                                         struct kobj_attribute *attr, char *buf)
1523 {
1524         struct hstate *h;
1525         unsigned long surplus_huge_pages;
1526         int nid;
1527
1528         h = kobj_to_hstate(kobj, &nid);
1529         if (nid == NUMA_NO_NODE)
1530                 surplus_huge_pages = h->surplus_huge_pages;
1531         else
1532                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1533
1534         return sprintf(buf, "%lu\n", surplus_huge_pages);
1535 }
1536 HSTATE_ATTR_RO(surplus_hugepages);
1537
1538 static struct attribute *hstate_attrs[] = {
1539         &nr_hugepages_attr.attr,
1540         &nr_overcommit_hugepages_attr.attr,
1541         &free_hugepages_attr.attr,
1542         &resv_hugepages_attr.attr,
1543         &surplus_hugepages_attr.attr,
1544 #ifdef CONFIG_NUMA
1545         &nr_hugepages_mempolicy_attr.attr,
1546 #endif
1547         NULL,
1548 };
1549
1550 static struct attribute_group hstate_attr_group = {
1551         .attrs = hstate_attrs,
1552 };
1553
1554 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1555                                     struct kobject **hstate_kobjs,
1556                                     struct attribute_group *hstate_attr_group)
1557 {
1558         int retval;
1559         int hi = h - hstates;
1560
1561         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1562         if (!hstate_kobjs[hi])
1563                 return -ENOMEM;
1564
1565         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1566         if (retval)
1567                 kobject_put(hstate_kobjs[hi]);
1568
1569         return retval;
1570 }
1571
1572 static void __init hugetlb_sysfs_init(void)
1573 {
1574         struct hstate *h;
1575         int err;
1576
1577         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1578         if (!hugepages_kobj)
1579                 return;
1580
1581         for_each_hstate(h) {
1582                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1583                                          hstate_kobjs, &hstate_attr_group);
1584                 if (err)
1585                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1586                                                                 h->name);
1587         }
1588 }
1589
1590 #ifdef CONFIG_NUMA
1591
1592 /*
1593  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1594  * with node sysdevs in node_devices[] using a parallel array.  The array
1595  * index of a node sysdev or _hstate == node id.
1596  * This is here to avoid any static dependency of the node sysdev driver, in
1597  * the base kernel, on the hugetlb module.
1598  */
1599 struct node_hstate {
1600         struct kobject          *hugepages_kobj;
1601         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1602 };
1603 struct node_hstate node_hstates[MAX_NUMNODES];
1604
1605 /*
1606  * A subset of global hstate attributes for node sysdevs
1607  */
1608 static struct attribute *per_node_hstate_attrs[] = {
1609         &nr_hugepages_attr.attr,
1610         &free_hugepages_attr.attr,
1611         &surplus_hugepages_attr.attr,
1612         NULL,
1613 };
1614
1615 static struct attribute_group per_node_hstate_attr_group = {
1616         .attrs = per_node_hstate_attrs,
1617 };
1618
1619 /*
1620  * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1621  * Returns node id via non-NULL nidp.
1622  */
1623 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1624 {
1625         int nid;
1626
1627         for (nid = 0; nid < nr_node_ids; nid++) {
1628                 struct node_hstate *nhs = &node_hstates[nid];
1629                 int i;
1630                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1631                         if (nhs->hstate_kobjs[i] == kobj) {
1632                                 if (nidp)
1633                                         *nidp = nid;
1634                                 return &hstates[i];
1635                         }
1636         }
1637
1638         BUG();
1639         return NULL;
1640 }
1641
1642 /*
1643  * Unregister hstate attributes from a single node sysdev.
1644  * No-op if no hstate attributes attached.
1645  */
1646 void hugetlb_unregister_node(struct node *node)
1647 {
1648         struct hstate *h;
1649         struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1650
1651         if (!nhs->hugepages_kobj)
1652                 return;         /* no hstate attributes */
1653
1654         for_each_hstate(h)
1655                 if (nhs->hstate_kobjs[h - hstates]) {
1656                         kobject_put(nhs->hstate_kobjs[h - hstates]);
1657                         nhs->hstate_kobjs[h - hstates] = NULL;
1658                 }
1659
1660         kobject_put(nhs->hugepages_kobj);
1661         nhs->hugepages_kobj = NULL;
1662 }
1663
1664 /*
1665  * hugetlb module exit:  unregister hstate attributes from node sysdevs
1666  * that have them.
1667  */
1668 static void hugetlb_unregister_all_nodes(void)
1669 {
1670         int nid;
1671
1672         /*
1673          * disable node sysdev registrations.
1674          */
1675         register_hugetlbfs_with_node(NULL, NULL);
1676
1677         /*
1678          * remove hstate attributes from any nodes that have them.
1679          */
1680         for (nid = 0; nid < nr_node_ids; nid++)
1681                 hugetlb_unregister_node(&node_devices[nid]);
1682 }
1683
1684 /*
1685  * Register hstate attributes for a single node sysdev.
1686  * No-op if attributes already registered.
1687  */
1688 void hugetlb_register_node(struct node *node)
1689 {
1690         struct hstate *h;
1691         struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1692         int err;
1693
1694         if (nhs->hugepages_kobj)
1695                 return;         /* already allocated */
1696
1697         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1698                                                         &node->sysdev.kobj);
1699         if (!nhs->hugepages_kobj)
1700                 return;
1701
1702         for_each_hstate(h) {
1703                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1704                                                 nhs->hstate_kobjs,
1705                                                 &per_node_hstate_attr_group);
1706                 if (err) {
1707                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1708                                         " for node %d\n",
1709                                                 h->name, node->sysdev.id);
1710                         hugetlb_unregister_node(node);
1711                         break;
1712                 }
1713         }
1714 }
1715
1716 /*
1717  * hugetlb init time:  register hstate attributes for all registered node
1718  * sysdevs of nodes that have memory.  All on-line nodes should have
1719  * registered their associated sysdev by this time.
1720  */
1721 static void hugetlb_register_all_nodes(void)
1722 {
1723         int nid;
1724
1725         for_each_node_state(nid, N_HIGH_MEMORY) {
1726                 struct node *node = &node_devices[nid];
1727                 if (node->sysdev.id == nid)
1728                         hugetlb_register_node(node);
1729         }
1730
1731         /*
1732          * Let the node sysdev driver know we're here so it can
1733          * [un]register hstate attributes on node hotplug.
1734          */
1735         register_hugetlbfs_with_node(hugetlb_register_node,
1736                                      hugetlb_unregister_node);
1737 }
1738 #else   /* !CONFIG_NUMA */
1739
1740 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1741 {
1742         BUG();
1743         if (nidp)
1744                 *nidp = -1;
1745         return NULL;
1746 }
1747
1748 static void hugetlb_unregister_all_nodes(void) { }
1749
1750 static void hugetlb_register_all_nodes(void) { }
1751
1752 #endif
1753
1754 static void __exit hugetlb_exit(void)
1755 {
1756         struct hstate *h;
1757
1758         hugetlb_unregister_all_nodes();
1759
1760         for_each_hstate(h) {
1761                 kobject_put(hstate_kobjs[h - hstates]);
1762         }
1763
1764         kobject_put(hugepages_kobj);
1765 }
1766 module_exit(hugetlb_exit);
1767
1768 static int __init hugetlb_init(void)
1769 {
1770         /* Some platform decide whether they support huge pages at boot
1771          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1772          * there is no such support
1773          */
1774         if (HPAGE_SHIFT == 0)
1775                 return 0;
1776
1777         if (!size_to_hstate(default_hstate_size)) {
1778                 default_hstate_size = HPAGE_SIZE;
1779                 if (!size_to_hstate(default_hstate_size))
1780                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1781         }
1782         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1783         if (default_hstate_max_huge_pages)
1784                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1785
1786         hugetlb_init_hstates();
1787
1788         gather_bootmem_prealloc();
1789
1790         report_hugepages();
1791
1792         hugetlb_sysfs_init();
1793
1794         hugetlb_register_all_nodes();
1795
1796         return 0;
1797 }
1798 module_init(hugetlb_init);
1799
1800 /* Should be called on processing a hugepagesz=... option */
1801 void __init hugetlb_add_hstate(unsigned order)
1802 {
1803         struct hstate *h;
1804         unsigned long i;
1805
1806         if (size_to_hstate(PAGE_SIZE << order)) {
1807                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1808                 return;
1809         }
1810         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1811         BUG_ON(order == 0);
1812         h = &hstates[max_hstate++];
1813         h->order = order;
1814         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1815         h->nr_huge_pages = 0;
1816         h->free_huge_pages = 0;
1817         for (i = 0; i < MAX_NUMNODES; ++i)
1818                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1819         h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1820         h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1821         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1822                                         huge_page_size(h)/1024);
1823
1824         parsed_hstate = h;
1825 }
1826
1827 static int __init hugetlb_nrpages_setup(char *s)
1828 {
1829         unsigned long *mhp;
1830         static unsigned long *last_mhp;
1831
1832         /*
1833          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1834          * so this hugepages= parameter goes to the "default hstate".
1835          */
1836         if (!max_hstate)
1837                 mhp = &default_hstate_max_huge_pages;
1838         else
1839                 mhp = &parsed_hstate->max_huge_pages;
1840
1841         if (mhp == last_mhp) {
1842                 printk(KERN_WARNING "hugepages= specified twice without "
1843                         "interleaving hugepagesz=, ignoring\n");
1844                 return 1;
1845         }
1846
1847         if (sscanf(s, "%lu", mhp) <= 0)
1848                 *mhp = 0;
1849
1850         /*
1851          * Global state is always initialized later in hugetlb_init.
1852          * But we need to allocate >= MAX_ORDER hstates here early to still
1853          * use the bootmem allocator.
1854          */
1855         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1856                 hugetlb_hstate_alloc_pages(parsed_hstate);
1857
1858         last_mhp = mhp;
1859
1860         return 1;
1861 }
1862 __setup("hugepages=", hugetlb_nrpages_setup);
1863
1864 static int __init hugetlb_default_setup(char *s)
1865 {
1866         default_hstate_size = memparse(s, &s);
1867         return 1;
1868 }
1869 __setup("default_hugepagesz=", hugetlb_default_setup);
1870
1871 static unsigned int cpuset_mems_nr(unsigned int *array)
1872 {
1873         int node;
1874         unsigned int nr = 0;
1875
1876         for_each_node_mask(node, cpuset_current_mems_allowed)
1877                 nr += array[node];
1878
1879         return nr;
1880 }
1881
1882 #ifdef CONFIG_SYSCTL
1883 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1884                          struct ctl_table *table, int write,
1885                          void __user *buffer, size_t *length, loff_t *ppos)
1886 {
1887         struct hstate *h = &default_hstate;
1888         unsigned long tmp;
1889         int ret;
1890
1891         tmp = h->max_huge_pages;
1892
1893         if (write && h->order >= MAX_ORDER)
1894                 return -EINVAL;
1895
1896         table->data = &tmp;
1897         table->maxlen = sizeof(unsigned long);
1898         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1899         if (ret)
1900                 goto out;
1901
1902         if (write) {
1903                 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1904                                                 GFP_KERNEL | __GFP_NORETRY);
1905                 if (!(obey_mempolicy &&
1906                                init_nodemask_of_mempolicy(nodes_allowed))) {
1907                         NODEMASK_FREE(nodes_allowed);
1908                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1909                 }
1910                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1911
1912                 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1913                         NODEMASK_FREE(nodes_allowed);
1914         }
1915 out:
1916         return ret;
1917 }
1918
1919 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1920                           void __user *buffer, size_t *length, loff_t *ppos)
1921 {
1922
1923         return hugetlb_sysctl_handler_common(false, table, write,
1924                                                         buffer, length, ppos);
1925 }
1926
1927 #ifdef CONFIG_NUMA
1928 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1929                           void __user *buffer, size_t *length, loff_t *ppos)
1930 {
1931         return hugetlb_sysctl_handler_common(true, table, write,
1932                                                         buffer, length, ppos);
1933 }
1934 #endif /* CONFIG_NUMA */
1935
1936 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1937                         void __user *buffer,
1938                         size_t *length, loff_t *ppos)
1939 {
1940         proc_dointvec(table, write, buffer, length, ppos);
1941         if (hugepages_treat_as_movable)
1942                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1943         else
1944                 htlb_alloc_mask = GFP_HIGHUSER;
1945         return 0;
1946 }
1947
1948 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1949                         void __user *buffer,
1950                         size_t *length, loff_t *ppos)
1951 {
1952         struct hstate *h = &default_hstate;
1953         unsigned long tmp;
1954         int ret;
1955
1956         tmp = h->nr_overcommit_huge_pages;
1957
1958         if (write && h->order >= MAX_ORDER)
1959                 return -EINVAL;
1960
1961         table->data = &tmp;
1962         table->maxlen = sizeof(unsigned long);
1963         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1964         if (ret)
1965                 goto out;
1966
1967         if (write) {
1968                 spin_lock(&hugetlb_lock);
1969                 h->nr_overcommit_huge_pages = tmp;
1970                 spin_unlock(&hugetlb_lock);
1971         }
1972 out:
1973         return ret;
1974 }
1975
1976 #endif /* CONFIG_SYSCTL */
1977
1978 void hugetlb_report_meminfo(struct seq_file *m)
1979 {
1980         struct hstate *h = &default_hstate;
1981         seq_printf(m,
1982                         "HugePages_Total:   %5lu\n"
1983                         "HugePages_Free:    %5lu\n"
1984                         "HugePages_Rsvd:    %5lu\n"
1985                         "HugePages_Surp:    %5lu\n"
1986                         "Hugepagesize:   %8lu kB\n",
1987                         h->nr_huge_pages,
1988                         h->free_huge_pages,
1989                         h->resv_huge_pages,
1990                         h->surplus_huge_pages,
1991                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1992 }
1993
1994 int hugetlb_report_node_meminfo(int nid, char *buf)
1995 {
1996         struct hstate *h = &default_hstate;
1997         return sprintf(buf,
1998                 "Node %d HugePages_Total: %5u\n"
1999                 "Node %d HugePages_Free:  %5u\n"
2000                 "Node %d HugePages_Surp:  %5u\n",
2001                 nid, h->nr_huge_pages_node[nid],
2002                 nid, h->free_huge_pages_node[nid],
2003                 nid, h->surplus_huge_pages_node[nid]);
2004 }
2005
2006 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2007 unsigned long hugetlb_total_pages(void)
2008 {
2009         struct hstate *h = &default_hstate;
2010         return h->nr_huge_pages * pages_per_huge_page(h);
2011 }
2012
2013 static int hugetlb_acct_memory(struct hstate *h, long delta)
2014 {
2015         int ret = -ENOMEM;
2016
2017         spin_lock(&hugetlb_lock);
2018         /*
2019          * When cpuset is configured, it breaks the strict hugetlb page
2020          * reservation as the accounting is done on a global variable. Such
2021          * reservation is completely rubbish in the presence of cpuset because
2022          * the reservation is not checked against page availability for the
2023          * current cpuset. Application can still potentially OOM'ed by kernel
2024          * with lack of free htlb page in cpuset that the task is in.
2025          * Attempt to enforce strict accounting with cpuset is almost
2026          * impossible (or too ugly) because cpuset is too fluid that
2027          * task or memory node can be dynamically moved between cpusets.
2028          *
2029          * The change of semantics for shared hugetlb mapping with cpuset is
2030          * undesirable. However, in order to preserve some of the semantics,
2031          * we fall back to check against current free page availability as
2032          * a best attempt and hopefully to minimize the impact of changing
2033          * semantics that cpuset has.
2034          */
2035         if (delta > 0) {
2036                 if (gather_surplus_pages(h, delta) < 0)
2037                         goto out;
2038
2039                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2040                         return_unused_surplus_pages(h, delta);
2041                         goto out;
2042                 }
2043         }
2044
2045         ret = 0;
2046         if (delta < 0)
2047                 return_unused_surplus_pages(h, (unsigned long) -delta);
2048
2049 out:
2050         spin_unlock(&hugetlb_lock);
2051         return ret;
2052 }
2053
2054 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2055 {
2056         struct resv_map *reservations = vma_resv_map(vma);
2057
2058         /*
2059          * This new VMA should share its siblings reservation map if present.
2060          * The VMA will only ever have a valid reservation map pointer where
2061          * it is being copied for another still existing VMA.  As that VMA
2062          * has a reference to the reservation map it cannot disappear until
2063          * after this open call completes.  It is therefore safe to take a
2064          * new reference here without additional locking.
2065          */
2066         if (reservations)
2067                 kref_get(&reservations->refs);
2068 }
2069
2070 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2071 {
2072         struct hstate *h = hstate_vma(vma);
2073         struct resv_map *reservations = vma_resv_map(vma);
2074         unsigned long reserve;
2075         unsigned long start;
2076         unsigned long end;
2077
2078         if (reservations) {
2079                 start = vma_hugecache_offset(h, vma, vma->vm_start);
2080                 end = vma_hugecache_offset(h, vma, vma->vm_end);
2081
2082                 reserve = (end - start) -
2083                         region_count(&reservations->regions, start, end);
2084
2085                 kref_put(&reservations->refs, resv_map_release);
2086
2087                 if (reserve) {
2088                         hugetlb_acct_memory(h, -reserve);
2089                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2090                 }
2091         }
2092 }
2093
2094 /*
2095  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2096  * handle_mm_fault() to try to instantiate regular-sized pages in the
2097  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2098  * this far.
2099  */
2100 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2101 {
2102         BUG();
2103         return 0;
2104 }
2105
2106 const struct vm_operations_struct hugetlb_vm_ops = {
2107         .fault = hugetlb_vm_op_fault,
2108         .open = hugetlb_vm_op_open,
2109         .close = hugetlb_vm_op_close,
2110 };
2111
2112 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2113                                 int writable)
2114 {
2115         pte_t entry;
2116
2117         if (writable) {
2118                 entry =
2119                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2120         } else {
2121                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2122         }
2123         entry = pte_mkyoung(entry);
2124         entry = pte_mkhuge(entry);
2125
2126         return entry;
2127 }
2128
2129 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2130                                    unsigned long address, pte_t *ptep)
2131 {
2132         pte_t entry;
2133
2134         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2135         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2136                 update_mmu_cache(vma, address, ptep);
2137 }
2138
2139
2140 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2141                             struct vm_area_struct *vma)
2142 {
2143         pte_t *src_pte, *dst_pte, entry;
2144         struct page *ptepage;
2145         unsigned long addr;
2146         int cow;
2147         struct hstate *h = hstate_vma(vma);
2148         unsigned long sz = huge_page_size(h);
2149
2150         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2151
2152         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2153                 src_pte = huge_pte_offset(src, addr);
2154                 if (!src_pte)
2155                         continue;
2156                 dst_pte = huge_pte_alloc(dst, addr, sz);
2157                 if (!dst_pte)
2158                         goto nomem;
2159
2160                 /* If the pagetables are shared don't copy or take references */
2161                 if (dst_pte == src_pte)
2162                         continue;
2163
2164                 spin_lock(&dst->page_table_lock);
2165                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2166                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2167                         if (cow)
2168                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2169                         entry = huge_ptep_get(src_pte);
2170                         ptepage = pte_page(entry);
2171                         get_page(ptepage);
2172                         page_dup_rmap(ptepage);
2173                         set_huge_pte_at(dst, addr, dst_pte, entry);
2174                 }
2175                 spin_unlock(&src->page_table_lock);
2176                 spin_unlock(&dst->page_table_lock);
2177         }
2178         return 0;
2179
2180 nomem:
2181         return -ENOMEM;
2182 }
2183
2184 static int is_hugetlb_entry_migration(pte_t pte)
2185 {
2186         swp_entry_t swp;
2187
2188         if (huge_pte_none(pte) || pte_present(pte))
2189                 return 0;
2190         swp = pte_to_swp_entry(pte);
2191         if (non_swap_entry(swp) && is_migration_entry(swp))
2192                 return 1;
2193         else
2194                 return 0;
2195 }
2196
2197 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2198 {
2199         swp_entry_t swp;
2200
2201         if (huge_pte_none(pte) || pte_present(pte))
2202                 return 0;
2203         swp = pte_to_swp_entry(pte);
2204         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2205                 return 1;
2206         else
2207                 return 0;
2208 }
2209
2210 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2211                             unsigned long end, struct page *ref_page)
2212 {
2213         struct mm_struct *mm = vma->vm_mm;
2214         unsigned long address;
2215         pte_t *ptep;
2216         pte_t pte;
2217         struct page *page;
2218         struct page *tmp;
2219         struct hstate *h = hstate_vma(vma);
2220         unsigned long sz = huge_page_size(h);
2221
2222         /*
2223          * A page gathering list, protected by per file i_mmap_mutex. The
2224          * lock is used to avoid list corruption from multiple unmapping
2225          * of the same page since we are using page->lru.
2226          */
2227         LIST_HEAD(page_list);
2228
2229         WARN_ON(!is_vm_hugetlb_page(vma));
2230         BUG_ON(start & ~huge_page_mask(h));
2231         BUG_ON(end & ~huge_page_mask(h));
2232
2233         mmu_notifier_invalidate_range_start(mm, start, end);
2234         spin_lock(&mm->page_table_lock);
2235         for (address = start; address < end; address += sz) {
2236                 ptep = huge_pte_offset(mm, address);
2237                 if (!ptep)
2238                         continue;
2239
2240                 if (huge_pmd_unshare(mm, &address, ptep))
2241                         continue;
2242
2243                 /*
2244                  * If a reference page is supplied, it is because a specific
2245                  * page is being unmapped, not a range. Ensure the page we
2246                  * are about to unmap is the actual page of interest.
2247                  */
2248                 if (ref_page) {
2249                         pte = huge_ptep_get(ptep);
2250                         if (huge_pte_none(pte))
2251                                 continue;
2252                         page = pte_page(pte);
2253                         if (page != ref_page)
2254                                 continue;
2255
2256                         /*
2257                          * Mark the VMA as having unmapped its page so that
2258                          * future faults in this VMA will fail rather than
2259                          * looking like data was lost
2260                          */
2261                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2262                 }
2263
2264                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2265                 if (huge_pte_none(pte))
2266                         continue;
2267
2268                 /*
2269                  * HWPoisoned hugepage is already unmapped and dropped reference
2270                  */
2271                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2272                         continue;
2273
2274                 page = pte_page(pte);
2275                 if (pte_dirty(pte))
2276                         set_page_dirty(page);
2277                 list_add(&page->lru, &page_list);
2278         }
2279         spin_unlock(&mm->page_table_lock);
2280         flush_tlb_range(vma, start, end);
2281         mmu_notifier_invalidate_range_end(mm, start, end);
2282         list_for_each_entry_safe(page, tmp, &page_list, lru) {
2283                 page_remove_rmap(page);
2284                 list_del(&page->lru);
2285                 put_page(page);
2286         }
2287 }
2288
2289 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2290                           unsigned long end, struct page *ref_page)
2291 {
2292         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2293         __unmap_hugepage_range(vma, start, end, ref_page);
2294         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2295 }
2296
2297 /*
2298  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2299  * mappping it owns the reserve page for. The intention is to unmap the page
2300  * from other VMAs and let the children be SIGKILLed if they are faulting the
2301  * same region.
2302  */
2303 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2304                                 struct page *page, unsigned long address)
2305 {
2306         struct hstate *h = hstate_vma(vma);
2307         struct vm_area_struct *iter_vma;
2308         struct address_space *mapping;
2309         struct prio_tree_iter iter;
2310         pgoff_t pgoff;
2311
2312         /*
2313          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2314          * from page cache lookup which is in HPAGE_SIZE units.
2315          */
2316         address = address & huge_page_mask(h);
2317         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2318                 + (vma->vm_pgoff >> PAGE_SHIFT);
2319         mapping = (struct address_space *)page_private(page);
2320
2321         /*
2322          * Take the mapping lock for the duration of the table walk. As
2323          * this mapping should be shared between all the VMAs,
2324          * __unmap_hugepage_range() is called as the lock is already held
2325          */
2326         mutex_lock(&mapping->i_mmap_mutex);
2327         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2328                 /* Do not unmap the current VMA */
2329                 if (iter_vma == vma)
2330                         continue;
2331
2332                 /*
2333                  * Unmap the page from other VMAs without their own reserves.
2334                  * They get marked to be SIGKILLed if they fault in these
2335                  * areas. This is because a future no-page fault on this VMA
2336                  * could insert a zeroed page instead of the data existing
2337                  * from the time of fork. This would look like data corruption
2338                  */
2339                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2340                         __unmap_hugepage_range(iter_vma,
2341                                 address, address + huge_page_size(h),
2342                                 page);
2343         }
2344         mutex_unlock(&mapping->i_mmap_mutex);
2345
2346         return 1;
2347 }
2348
2349 /*
2350  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2351  */
2352 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2353                         unsigned long address, pte_t *ptep, pte_t pte,
2354                         struct page *pagecache_page)
2355 {
2356         struct hstate *h = hstate_vma(vma);
2357         struct page *old_page, *new_page;
2358         int avoidcopy;
2359         int outside_reserve = 0;
2360
2361         old_page = pte_page(pte);
2362
2363 retry_avoidcopy:
2364         /* If no-one else is actually using this page, avoid the copy
2365          * and just make the page writable */
2366         avoidcopy = (page_mapcount(old_page) == 1);
2367         if (avoidcopy) {
2368                 if (PageAnon(old_page))
2369                         page_move_anon_rmap(old_page, vma, address);
2370                 set_huge_ptep_writable(vma, address, ptep);
2371                 return 0;
2372         }
2373
2374         /*
2375          * If the process that created a MAP_PRIVATE mapping is about to
2376          * perform a COW due to a shared page count, attempt to satisfy
2377          * the allocation without using the existing reserves. The pagecache
2378          * page is used to determine if the reserve at this address was
2379          * consumed or not. If reserves were used, a partial faulted mapping
2380          * at the time of fork() could consume its reserves on COW instead
2381          * of the full address range.
2382          */
2383         if (!(vma->vm_flags & VM_MAYSHARE) &&
2384                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2385                         old_page != pagecache_page)
2386                 outside_reserve = 1;
2387
2388         page_cache_get(old_page);
2389
2390         /* Drop page_table_lock as buddy allocator may be called */
2391         spin_unlock(&mm->page_table_lock);
2392         new_page = alloc_huge_page(vma, address, outside_reserve);
2393
2394         if (IS_ERR(new_page)) {
2395                 page_cache_release(old_page);
2396
2397                 /*
2398                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2399                  * it is due to references held by a child and an insufficient
2400                  * huge page pool. To guarantee the original mappers
2401                  * reliability, unmap the page from child processes. The child
2402                  * may get SIGKILLed if it later faults.
2403                  */
2404                 if (outside_reserve) {
2405                         BUG_ON(huge_pte_none(pte));
2406                         if (unmap_ref_private(mm, vma, old_page, address)) {
2407                                 BUG_ON(page_count(old_page) != 1);
2408                                 BUG_ON(huge_pte_none(pte));
2409                                 spin_lock(&mm->page_table_lock);
2410                                 goto retry_avoidcopy;
2411                         }
2412                         WARN_ON_ONCE(1);
2413                 }
2414
2415                 /* Caller expects lock to be held */
2416                 spin_lock(&mm->page_table_lock);
2417                 return -PTR_ERR(new_page);
2418         }
2419
2420         /*
2421          * When the original hugepage is shared one, it does not have
2422          * anon_vma prepared.
2423          */
2424         if (unlikely(anon_vma_prepare(vma))) {
2425                 /* Caller expects lock to be held */
2426                 spin_lock(&mm->page_table_lock);
2427                 return VM_FAULT_OOM;
2428         }
2429
2430         copy_user_huge_page(new_page, old_page, address, vma,
2431                             pages_per_huge_page(h));
2432         __SetPageUptodate(new_page);
2433
2434         /*
2435          * Retake the page_table_lock to check for racing updates
2436          * before the page tables are altered
2437          */
2438         spin_lock(&mm->page_table_lock);
2439         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2440         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2441                 /* Break COW */
2442                 mmu_notifier_invalidate_range_start(mm,
2443                         address & huge_page_mask(h),
2444                         (address & huge_page_mask(h)) + huge_page_size(h));
2445                 huge_ptep_clear_flush(vma, address, ptep);
2446                 set_huge_pte_at(mm, address, ptep,
2447                                 make_huge_pte(vma, new_page, 1));
2448                 page_remove_rmap(old_page);
2449                 hugepage_add_new_anon_rmap(new_page, vma, address);
2450                 /* Make the old page be freed below */
2451                 new_page = old_page;
2452                 mmu_notifier_invalidate_range_end(mm,
2453                         address & huge_page_mask(h),
2454                         (address & huge_page_mask(h)) + huge_page_size(h));
2455         }
2456         page_cache_release(new_page);
2457         page_cache_release(old_page);
2458         return 0;
2459 }
2460
2461 /* Return the pagecache page at a given address within a VMA */
2462 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2463                         struct vm_area_struct *vma, unsigned long address)
2464 {
2465         struct address_space *mapping;
2466         pgoff_t idx;
2467
2468         mapping = vma->vm_file->f_mapping;
2469         idx = vma_hugecache_offset(h, vma, address);
2470
2471         return find_lock_page(mapping, idx);
2472 }
2473
2474 /*
2475  * Return whether there is a pagecache page to back given address within VMA.
2476  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2477  */
2478 static bool hugetlbfs_pagecache_present(struct hstate *h,
2479                         struct vm_area_struct *vma, unsigned long address)
2480 {
2481         struct address_space *mapping;
2482         pgoff_t idx;
2483         struct page *page;
2484
2485         mapping = vma->vm_file->f_mapping;
2486         idx = vma_hugecache_offset(h, vma, address);
2487
2488         page = find_get_page(mapping, idx);
2489         if (page)
2490                 put_page(page);
2491         return page != NULL;
2492 }
2493
2494 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2495                         unsigned long address, pte_t *ptep, unsigned int flags)
2496 {
2497         struct hstate *h = hstate_vma(vma);
2498         int ret = VM_FAULT_SIGBUS;
2499         pgoff_t idx;
2500         unsigned long size;
2501         struct page *page;
2502         struct address_space *mapping;
2503         pte_t new_pte;
2504
2505         /*
2506          * Currently, we are forced to kill the process in the event the
2507          * original mapper has unmapped pages from the child due to a failed
2508          * COW. Warn that such a situation has occurred as it may not be obvious
2509          */
2510         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2511                 printk(KERN_WARNING
2512                         "PID %d killed due to inadequate hugepage pool\n",
2513                         current->pid);
2514                 return ret;
2515         }
2516
2517         mapping = vma->vm_file->f_mapping;
2518         idx = vma_hugecache_offset(h, vma, address);
2519
2520         /*
2521          * Use page lock to guard against racing truncation
2522          * before we get page_table_lock.
2523          */
2524 retry:
2525         page = find_lock_page(mapping, idx);
2526         if (!page) {
2527                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2528                 if (idx >= size)
2529                         goto out;
2530                 page = alloc_huge_page(vma, address, 0);
2531                 if (IS_ERR(page)) {
2532                         ret = -PTR_ERR(page);
2533                         goto out;
2534                 }
2535                 clear_huge_page(page, address, pages_per_huge_page(h));
2536                 __SetPageUptodate(page);
2537
2538                 if (vma->vm_flags & VM_MAYSHARE) {
2539                         int err;
2540                         struct inode *inode = mapping->host;
2541
2542                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2543                         if (err) {
2544                                 put_page(page);
2545                                 if (err == -EEXIST)
2546                                         goto retry;
2547                                 goto out;
2548                         }
2549
2550                         spin_lock(&inode->i_lock);
2551                         inode->i_blocks += blocks_per_huge_page(h);
2552                         spin_unlock(&inode->i_lock);
2553                         page_dup_rmap(page);
2554                 } else {
2555                         lock_page(page);
2556                         if (unlikely(anon_vma_prepare(vma))) {
2557                                 ret = VM_FAULT_OOM;
2558                                 goto backout_unlocked;
2559                         }
2560                         hugepage_add_new_anon_rmap(page, vma, address);
2561                 }
2562         } else {
2563                 /*
2564                  * If memory error occurs between mmap() and fault, some process
2565                  * don't have hwpoisoned swap entry for errored virtual address.
2566                  * So we need to block hugepage fault by PG_hwpoison bit check.
2567                  */
2568                 if (unlikely(PageHWPoison(page))) {
2569                         ret = VM_FAULT_HWPOISON |
2570                               VM_FAULT_SET_HINDEX(h - hstates);
2571                         goto backout_unlocked;
2572                 }
2573                 page_dup_rmap(page);
2574         }
2575
2576         /*
2577          * If we are going to COW a private mapping later, we examine the
2578          * pending reservations for this page now. This will ensure that
2579          * any allocations necessary to record that reservation occur outside
2580          * the spinlock.
2581          */
2582         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2583                 if (vma_needs_reservation(h, vma, address) < 0) {
2584                         ret = VM_FAULT_OOM;
2585                         goto backout_unlocked;
2586                 }
2587
2588         spin_lock(&mm->page_table_lock);
2589         size = i_size_read(mapping->host) >> huge_page_shift(h);
2590         if (idx >= size)
2591                 goto backout;
2592
2593         ret = 0;
2594         if (!huge_pte_none(huge_ptep_get(ptep)))
2595                 goto backout;
2596
2597         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2598                                 && (vma->vm_flags & VM_SHARED)));
2599         set_huge_pte_at(mm, address, ptep, new_pte);
2600
2601         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2602                 /* Optimization, do the COW without a second fault */
2603                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2604         }
2605
2606         spin_unlock(&mm->page_table_lock);
2607         unlock_page(page);
2608 out:
2609         return ret;
2610
2611 backout:
2612         spin_unlock(&mm->page_table_lock);
2613 backout_unlocked:
2614         unlock_page(page);
2615         put_page(page);
2616         goto out;
2617 }
2618
2619 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2620                         unsigned long address, unsigned int flags)
2621 {
2622         pte_t *ptep;
2623         pte_t entry;
2624         int ret;
2625         struct page *page = NULL;
2626         struct page *pagecache_page = NULL;
2627         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2628         struct hstate *h = hstate_vma(vma);
2629
2630         ptep = huge_pte_offset(mm, address);
2631         if (ptep) {
2632                 entry = huge_ptep_get(ptep);
2633                 if (unlikely(is_hugetlb_entry_migration(entry))) {
2634                         migration_entry_wait(mm, (pmd_t *)ptep, address);
2635                         return 0;
2636                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2637                         return VM_FAULT_HWPOISON_LARGE |
2638                                VM_FAULT_SET_HINDEX(h - hstates);
2639         }
2640
2641         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2642         if (!ptep)
2643                 return VM_FAULT_OOM;
2644
2645         /*
2646          * Serialize hugepage allocation and instantiation, so that we don't
2647          * get spurious allocation failures if two CPUs race to instantiate
2648          * the same page in the page cache.
2649          */
2650         mutex_lock(&hugetlb_instantiation_mutex);
2651         entry = huge_ptep_get(ptep);
2652         if (huge_pte_none(entry)) {
2653                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2654                 goto out_mutex;
2655         }
2656
2657         ret = 0;
2658
2659         /*
2660          * If we are going to COW the mapping later, we examine the pending
2661          * reservations for this page now. This will ensure that any
2662          * allocations necessary to record that reservation occur outside the
2663          * spinlock. For private mappings, we also lookup the pagecache
2664          * page now as it is used to determine if a reservation has been
2665          * consumed.
2666          */
2667         if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2668                 if (vma_needs_reservation(h, vma, address) < 0) {
2669                         ret = VM_FAULT_OOM;
2670                         goto out_mutex;
2671                 }
2672
2673                 if (!(vma->vm_flags & VM_MAYSHARE))
2674                         pagecache_page = hugetlbfs_pagecache_page(h,
2675                                                                 vma, address);
2676         }
2677
2678         /*
2679          * hugetlb_cow() requires page locks of pte_page(entry) and
2680          * pagecache_page, so here we need take the former one
2681          * when page != pagecache_page or !pagecache_page.
2682          * Note that locking order is always pagecache_page -> page,
2683          * so no worry about deadlock.
2684          */
2685         page = pte_page(entry);
2686         if (page != pagecache_page)
2687                 lock_page(page);
2688
2689         spin_lock(&mm->page_table_lock);
2690         /* Check for a racing update before calling hugetlb_cow */
2691         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2692                 goto out_page_table_lock;
2693
2694
2695         if (flags & FAULT_FLAG_WRITE) {
2696                 if (!pte_write(entry)) {
2697                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2698                                                         pagecache_page);
2699                         goto out_page_table_lock;
2700                 }
2701                 entry = pte_mkdirty(entry);
2702         }
2703         entry = pte_mkyoung(entry);
2704         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2705                                                 flags & FAULT_FLAG_WRITE))
2706                 update_mmu_cache(vma, address, ptep);
2707
2708 out_page_table_lock:
2709         spin_unlock(&mm->page_table_lock);
2710
2711         if (pagecache_page) {
2712                 unlock_page(pagecache_page);
2713                 put_page(pagecache_page);
2714         }
2715         if (page != pagecache_page)
2716                 unlock_page(page);
2717
2718 out_mutex:
2719         mutex_unlock(&hugetlb_instantiation_mutex);
2720
2721         return ret;
2722 }
2723
2724 /* Can be overriden by architectures */
2725 __attribute__((weak)) struct page *
2726 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2727                pud_t *pud, int write)
2728 {
2729         BUG();
2730         return NULL;
2731 }
2732
2733 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2734                         struct page **pages, struct vm_area_struct **vmas,
2735                         unsigned long *position, int *length, int i,
2736                         unsigned int flags)
2737 {
2738         unsigned long pfn_offset;
2739         unsigned long vaddr = *position;
2740         int remainder = *length;
2741         struct hstate *h = hstate_vma(vma);
2742
2743         spin_lock(&mm->page_table_lock);
2744         while (vaddr < vma->vm_end && remainder) {
2745                 pte_t *pte;
2746                 int absent;
2747                 struct page *page;
2748
2749                 /*
2750                  * Some archs (sparc64, sh*) have multiple pte_ts to
2751                  * each hugepage.  We have to make sure we get the
2752                  * first, for the page indexing below to work.
2753                  */
2754                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2755                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2756
2757                 /*
2758                  * When coredumping, it suits get_dump_page if we just return
2759                  * an error where there's an empty slot with no huge pagecache
2760                  * to back it.  This way, we avoid allocating a hugepage, and
2761                  * the sparse dumpfile avoids allocating disk blocks, but its
2762                  * huge holes still show up with zeroes where they need to be.
2763                  */
2764                 if (absent && (flags & FOLL_DUMP) &&
2765                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2766                         remainder = 0;
2767                         break;
2768                 }
2769
2770                 if (absent ||
2771                     ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2772                         int ret;
2773
2774                         spin_unlock(&mm->page_table_lock);
2775                         ret = hugetlb_fault(mm, vma, vaddr,
2776                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2777                         spin_lock(&mm->page_table_lock);
2778                         if (!(ret & VM_FAULT_ERROR))
2779                                 continue;
2780
2781                         remainder = 0;
2782                         break;
2783                 }
2784
2785                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2786                 page = pte_page(huge_ptep_get(pte));
2787 same_page:
2788                 if (pages) {
2789                         pages[i] = mem_map_offset(page, pfn_offset);
2790                         get_page(pages[i]);
2791                 }
2792
2793                 if (vmas)
2794                         vmas[i] = vma;
2795
2796                 vaddr += PAGE_SIZE;
2797                 ++pfn_offset;
2798                 --remainder;
2799                 ++i;
2800                 if (vaddr < vma->vm_end && remainder &&
2801                                 pfn_offset < pages_per_huge_page(h)) {
2802                         /*
2803                          * We use pfn_offset to avoid touching the pageframes
2804                          * of this compound page.
2805                          */
2806                         goto same_page;
2807                 }
2808         }
2809         spin_unlock(&mm->page_table_lock);
2810         *length = remainder;
2811         *position = vaddr;
2812
2813         return i ? i : -EFAULT;
2814 }
2815
2816 void hugetlb_change_protection(struct vm_area_struct *vma,
2817                 unsigned long address, unsigned long end, pgprot_t newprot)
2818 {
2819         struct mm_struct *mm = vma->vm_mm;
2820         unsigned long start = address;
2821         pte_t *ptep;
2822         pte_t pte;
2823         struct hstate *h = hstate_vma(vma);
2824
2825         BUG_ON(address >= end);
2826         flush_cache_range(vma, address, end);
2827
2828         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2829         spin_lock(&mm->page_table_lock);
2830         for (; address < end; address += huge_page_size(h)) {
2831                 ptep = huge_pte_offset(mm, address);
2832                 if (!ptep)
2833                         continue;
2834                 if (huge_pmd_unshare(mm, &address, ptep))
2835                         continue;
2836                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2837                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2838                         pte = pte_mkhuge(pte_modify(pte, newprot));
2839                         set_huge_pte_at(mm, address, ptep, pte);
2840                 }
2841         }
2842         spin_unlock(&mm->page_table_lock);
2843         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2844
2845         flush_tlb_range(vma, start, end);
2846 }
2847
2848 int hugetlb_reserve_pages(struct inode *inode,
2849                                         long from, long to,
2850                                         struct vm_area_struct *vma,
2851                                         vm_flags_t vm_flags)
2852 {
2853         long ret, chg;
2854         struct hstate *h = hstate_inode(inode);
2855
2856         /*
2857          * Only apply hugepage reservation if asked. At fault time, an
2858          * attempt will be made for VM_NORESERVE to allocate a page
2859          * and filesystem quota without using reserves
2860          */
2861         if (vm_flags & VM_NORESERVE)
2862                 return 0;
2863
2864         /*
2865          * Shared mappings base their reservation on the number of pages that
2866          * are already allocated on behalf of the file. Private mappings need
2867          * to reserve the full area even if read-only as mprotect() may be
2868          * called to make the mapping read-write. Assume !vma is a shm mapping
2869          */
2870         if (!vma || vma->vm_flags & VM_MAYSHARE)
2871                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2872         else {
2873                 struct resv_map *resv_map = resv_map_alloc();
2874                 if (!resv_map)
2875                         return -ENOMEM;
2876
2877                 chg = to - from;
2878
2879                 set_vma_resv_map(vma, resv_map);
2880                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2881         }
2882
2883         if (chg < 0)
2884                 return chg;
2885
2886         /* There must be enough filesystem quota for the mapping */
2887         if (hugetlb_get_quota(inode->i_mapping, chg))
2888                 return -ENOSPC;
2889
2890         /*
2891          * Check enough hugepages are available for the reservation.
2892          * Hand back the quota if there are not
2893          */
2894         ret = hugetlb_acct_memory(h, chg);
2895         if (ret < 0) {
2896                 hugetlb_put_quota(inode->i_mapping, chg);
2897                 return ret;
2898         }
2899
2900         /*
2901          * Account for the reservations made. Shared mappings record regions
2902          * that have reservations as they are shared by multiple VMAs.
2903          * When the last VMA disappears, the region map says how much
2904          * the reservation was and the page cache tells how much of
2905          * the reservation was consumed. Private mappings are per-VMA and
2906          * only the consumed reservations are tracked. When the VMA
2907          * disappears, the original reservation is the VMA size and the
2908          * consumed reservations are stored in the map. Hence, nothing
2909          * else has to be done for private mappings here
2910          */
2911         if (!vma || vma->vm_flags & VM_MAYSHARE)
2912                 region_add(&inode->i_mapping->private_list, from, to);
2913         return 0;
2914 }
2915
2916 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2917 {
2918         struct hstate *h = hstate_inode(inode);
2919         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2920
2921         spin_lock(&inode->i_lock);
2922         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2923         spin_unlock(&inode->i_lock);
2924
2925         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2926         hugetlb_acct_memory(h, -(chg - freed));
2927 }
2928
2929 #ifdef CONFIG_MEMORY_FAILURE
2930
2931 /* Should be called in hugetlb_lock */
2932 static int is_hugepage_on_freelist(struct page *hpage)
2933 {
2934         struct page *page;
2935         struct page *tmp;
2936         struct hstate *h = page_hstate(hpage);
2937         int nid = page_to_nid(hpage);
2938
2939         list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
2940                 if (page == hpage)
2941                         return 1;
2942         return 0;
2943 }
2944
2945 /*
2946  * This function is called from memory failure code.
2947  * Assume the caller holds page lock of the head page.
2948  */
2949 int dequeue_hwpoisoned_huge_page(struct page *hpage)
2950 {
2951         struct hstate *h = page_hstate(hpage);
2952         int nid = page_to_nid(hpage);
2953         int ret = -EBUSY;
2954
2955         spin_lock(&hugetlb_lock);
2956         if (is_hugepage_on_freelist(hpage)) {
2957                 list_del(&hpage->lru);
2958                 set_page_refcounted(hpage);
2959                 h->free_huge_pages--;
2960                 h->free_huge_pages_node[nid]--;
2961                 ret = 0;
2962         }
2963         spin_unlock(&hugetlb_lock);
2964         return ret;
2965 }
2966 #endif