ab89d6f382d1274a8294f8fd385f82d864400674
[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                 set_page_count(p, 0);
580                 p->first_page = page;
581         }
582 }
583
584 int PageHuge(struct page *page)
585 {
586         compound_page_dtor *dtor;
587
588         if (!PageCompound(page))
589                 return 0;
590
591         page = compound_head(page);
592         dtor = get_compound_page_dtor(page);
593
594         return dtor == free_huge_page;
595 }
596 EXPORT_SYMBOL_GPL(PageHuge);
597
598 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
599 {
600         struct page *page;
601
602         if (h->order >= MAX_ORDER)
603                 return NULL;
604
605         page = alloc_pages_exact_node(nid,
606                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
607                                                 __GFP_REPEAT|__GFP_NOWARN,
608                 huge_page_order(h));
609         if (page) {
610                 if (arch_prepare_hugepage(page)) {
611                         __free_pages(page, huge_page_order(h));
612                         return NULL;
613                 }
614                 prep_new_huge_page(h, page, nid);
615         }
616
617         return page;
618 }
619
620 /*
621  * common helper functions for hstate_next_node_to_{alloc|free}.
622  * We may have allocated or freed a huge page based on a different
623  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
624  * be outside of *nodes_allowed.  Ensure that we use an allowed
625  * node for alloc or free.
626  */
627 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
628 {
629         nid = next_node(nid, *nodes_allowed);
630         if (nid == MAX_NUMNODES)
631                 nid = first_node(*nodes_allowed);
632         VM_BUG_ON(nid >= MAX_NUMNODES);
633
634         return nid;
635 }
636
637 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
638 {
639         if (!node_isset(nid, *nodes_allowed))
640                 nid = next_node_allowed(nid, nodes_allowed);
641         return nid;
642 }
643
644 /*
645  * returns the previously saved node ["this node"] from which to
646  * allocate a persistent huge page for the pool and advance the
647  * next node from which to allocate, handling wrap at end of node
648  * mask.
649  */
650 static int hstate_next_node_to_alloc(struct hstate *h,
651                                         nodemask_t *nodes_allowed)
652 {
653         int nid;
654
655         VM_BUG_ON(!nodes_allowed);
656
657         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
658         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
659
660         return nid;
661 }
662
663 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
664 {
665         struct page *page;
666         int start_nid;
667         int next_nid;
668         int ret = 0;
669
670         start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
671         next_nid = start_nid;
672
673         do {
674                 page = alloc_fresh_huge_page_node(h, next_nid);
675                 if (page) {
676                         ret = 1;
677                         break;
678                 }
679                 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
680         } while (next_nid != start_nid);
681
682         if (ret)
683                 count_vm_event(HTLB_BUDDY_PGALLOC);
684         else
685                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
686
687         return ret;
688 }
689
690 /*
691  * helper for free_pool_huge_page() - return the previously saved
692  * node ["this node"] from which to free a huge page.  Advance the
693  * next node id whether or not we find a free huge page to free so
694  * that the next attempt to free addresses the next node.
695  */
696 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
697 {
698         int nid;
699
700         VM_BUG_ON(!nodes_allowed);
701
702         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
703         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
704
705         return nid;
706 }
707
708 /*
709  * Free huge page from pool from next node to free.
710  * Attempt to keep persistent huge pages more or less
711  * balanced over allowed nodes.
712  * Called with hugetlb_lock locked.
713  */
714 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
715                                                          bool acct_surplus)
716 {
717         int start_nid;
718         int next_nid;
719         int ret = 0;
720
721         start_nid = hstate_next_node_to_free(h, nodes_allowed);
722         next_nid = start_nid;
723
724         do {
725                 /*
726                  * If we're returning unused surplus pages, only examine
727                  * nodes with surplus pages.
728                  */
729                 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
730                     !list_empty(&h->hugepage_freelists[next_nid])) {
731                         struct page *page =
732                                 list_entry(h->hugepage_freelists[next_nid].next,
733                                           struct page, lru);
734                         list_del(&page->lru);
735                         h->free_huge_pages--;
736                         h->free_huge_pages_node[next_nid]--;
737                         if (acct_surplus) {
738                                 h->surplus_huge_pages--;
739                                 h->surplus_huge_pages_node[next_nid]--;
740                         }
741                         update_and_free_page(h, page);
742                         ret = 1;
743                         break;
744                 }
745                 next_nid = hstate_next_node_to_free(h, nodes_allowed);
746         } while (next_nid != start_nid);
747
748         return ret;
749 }
750
751 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
752 {
753         struct page *page;
754         unsigned int r_nid;
755
756         if (h->order >= MAX_ORDER)
757                 return NULL;
758
759         /*
760          * Assume we will successfully allocate the surplus page to
761          * prevent racing processes from causing the surplus to exceed
762          * overcommit
763          *
764          * This however introduces a different race, where a process B
765          * tries to grow the static hugepage pool while alloc_pages() is
766          * called by process A. B will only examine the per-node
767          * counters in determining if surplus huge pages can be
768          * converted to normal huge pages in adjust_pool_surplus(). A
769          * won't be able to increment the per-node counter, until the
770          * lock is dropped by B, but B doesn't drop hugetlb_lock until
771          * no more huge pages can be converted from surplus to normal
772          * state (and doesn't try to convert again). Thus, we have a
773          * case where a surplus huge page exists, the pool is grown, and
774          * the surplus huge page still exists after, even though it
775          * should just have been converted to a normal huge page. This
776          * does not leak memory, though, as the hugepage will be freed
777          * once it is out of use. It also does not allow the counters to
778          * go out of whack in adjust_pool_surplus() as we don't modify
779          * the node values until we've gotten the hugepage and only the
780          * per-node value is checked there.
781          */
782         spin_lock(&hugetlb_lock);
783         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
784                 spin_unlock(&hugetlb_lock);
785                 return NULL;
786         } else {
787                 h->nr_huge_pages++;
788                 h->surplus_huge_pages++;
789         }
790         spin_unlock(&hugetlb_lock);
791
792         if (nid == NUMA_NO_NODE)
793                 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
794                                    __GFP_REPEAT|__GFP_NOWARN,
795                                    huge_page_order(h));
796         else
797                 page = alloc_pages_exact_node(nid,
798                         htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
799                         __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
800
801         if (page && arch_prepare_hugepage(page)) {
802                 __free_pages(page, huge_page_order(h));
803                 return NULL;
804         }
805
806         spin_lock(&hugetlb_lock);
807         if (page) {
808                 r_nid = page_to_nid(page);
809                 set_compound_page_dtor(page, free_huge_page);
810                 /*
811                  * We incremented the global counters already
812                  */
813                 h->nr_huge_pages_node[r_nid]++;
814                 h->surplus_huge_pages_node[r_nid]++;
815                 __count_vm_event(HTLB_BUDDY_PGALLOC);
816         } else {
817                 h->nr_huge_pages--;
818                 h->surplus_huge_pages--;
819                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
820         }
821         spin_unlock(&hugetlb_lock);
822
823         return page;
824 }
825
826 /*
827  * This allocation function is useful in the context where vma is irrelevant.
828  * E.g. soft-offlining uses this function because it only cares physical
829  * address of error page.
830  */
831 struct page *alloc_huge_page_node(struct hstate *h, int nid)
832 {
833         struct page *page;
834
835         spin_lock(&hugetlb_lock);
836         page = dequeue_huge_page_node(h, nid);
837         spin_unlock(&hugetlb_lock);
838
839         if (!page)
840                 page = alloc_buddy_huge_page(h, nid);
841
842         return page;
843 }
844
845 /*
846  * Increase the hugetlb pool such that it can accommodate a reservation
847  * of size 'delta'.
848  */
849 static int gather_surplus_pages(struct hstate *h, int delta)
850 {
851         struct list_head surplus_list;
852         struct page *page, *tmp;
853         int ret, i;
854         int needed, allocated;
855
856         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
857         if (needed <= 0) {
858                 h->resv_huge_pages += delta;
859                 return 0;
860         }
861
862         allocated = 0;
863         INIT_LIST_HEAD(&surplus_list);
864
865         ret = -ENOMEM;
866 retry:
867         spin_unlock(&hugetlb_lock);
868         for (i = 0; i < needed; i++) {
869                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
870                 if (!page)
871                         /*
872                          * We were not able to allocate enough pages to
873                          * satisfy the entire reservation so we free what
874                          * we've allocated so far.
875                          */
876                         goto free;
877
878                 list_add(&page->lru, &surplus_list);
879         }
880         allocated += needed;
881
882         /*
883          * After retaking hugetlb_lock, we need to recalculate 'needed'
884          * because either resv_huge_pages or free_huge_pages may have changed.
885          */
886         spin_lock(&hugetlb_lock);
887         needed = (h->resv_huge_pages + delta) -
888                         (h->free_huge_pages + allocated);
889         if (needed > 0)
890                 goto retry;
891
892         /*
893          * The surplus_list now contains _at_least_ the number of extra pages
894          * needed to accommodate the reservation.  Add the appropriate number
895          * of pages to the hugetlb pool and free the extras back to the buddy
896          * allocator.  Commit the entire reservation here to prevent another
897          * process from stealing the pages as they are added to the pool but
898          * before they are reserved.
899          */
900         needed += allocated;
901         h->resv_huge_pages += delta;
902         ret = 0;
903
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         spin_unlock(&hugetlb_lock);
918
919         /* Free unnecessary surplus pages to the buddy allocator */
920 free:
921         if (!list_empty(&surplus_list)) {
922                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
923                         list_del(&page->lru);
924                         put_page(page);
925                 }
926         }
927         spin_lock(&hugetlb_lock);
928
929         return ret;
930 }
931
932 /*
933  * When releasing a hugetlb pool reservation, any surplus pages that were
934  * allocated to satisfy the reservation must be explicitly freed if they were
935  * never used.
936  * Called with hugetlb_lock held.
937  */
938 static void return_unused_surplus_pages(struct hstate *h,
939                                         unsigned long unused_resv_pages)
940 {
941         unsigned long nr_pages;
942
943         /* Uncommit the reservation */
944         h->resv_huge_pages -= unused_resv_pages;
945
946         /* Cannot return gigantic pages currently */
947         if (h->order >= MAX_ORDER)
948                 return;
949
950         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
951
952         /*
953          * We want to release as many surplus pages as possible, spread
954          * evenly across all nodes with memory. Iterate across these nodes
955          * until we can no longer free unreserved surplus pages. This occurs
956          * when the nodes with surplus pages have no free pages.
957          * free_pool_huge_page() will balance the the freed pages across the
958          * on-line nodes with memory and will handle the hstate accounting.
959          */
960         while (nr_pages--) {
961                 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
962                         break;
963         }
964 }
965
966 /*
967  * Determine if the huge page at addr within the vma has an associated
968  * reservation.  Where it does not we will need to logically increase
969  * reservation and actually increase quota before an allocation can occur.
970  * Where any new reservation would be required the reservation change is
971  * prepared, but not committed.  Once the page has been quota'd allocated
972  * an instantiated the change should be committed via vma_commit_reservation.
973  * No action is required on failure.
974  */
975 static long vma_needs_reservation(struct hstate *h,
976                         struct vm_area_struct *vma, unsigned long addr)
977 {
978         struct address_space *mapping = vma->vm_file->f_mapping;
979         struct inode *inode = mapping->host;
980
981         if (vma->vm_flags & VM_MAYSHARE) {
982                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
983                 return region_chg(&inode->i_mapping->private_list,
984                                                         idx, idx + 1);
985
986         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
987                 return 1;
988
989         } else  {
990                 long err;
991                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
992                 struct resv_map *reservations = vma_resv_map(vma);
993
994                 err = region_chg(&reservations->regions, idx, idx + 1);
995                 if (err < 0)
996                         return err;
997                 return 0;
998         }
999 }
1000 static void vma_commit_reservation(struct hstate *h,
1001                         struct vm_area_struct *vma, unsigned long addr)
1002 {
1003         struct address_space *mapping = vma->vm_file->f_mapping;
1004         struct inode *inode = mapping->host;
1005
1006         if (vma->vm_flags & VM_MAYSHARE) {
1007                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1008                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1009
1010         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1011                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1012                 struct resv_map *reservations = vma_resv_map(vma);
1013
1014                 /* Mark this page used in the map. */
1015                 region_add(&reservations->regions, idx, idx + 1);
1016         }
1017 }
1018
1019 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1020                                     unsigned long addr, int avoid_reserve)
1021 {
1022         struct hstate *h = hstate_vma(vma);
1023         struct page *page;
1024         struct address_space *mapping = vma->vm_file->f_mapping;
1025         struct inode *inode = mapping->host;
1026         long chg;
1027
1028         /*
1029          * Processes that did not create the mapping will have no reserves and
1030          * will not have accounted against quota. Check that the quota can be
1031          * made before satisfying the allocation
1032          * MAP_NORESERVE mappings may also need pages and quota allocated
1033          * if no reserve mapping overlaps.
1034          */
1035         chg = vma_needs_reservation(h, vma, addr);
1036         if (chg < 0)
1037                 return ERR_PTR(-VM_FAULT_OOM);
1038         if (chg)
1039                 if (hugetlb_get_quota(inode->i_mapping, chg))
1040                         return ERR_PTR(-VM_FAULT_SIGBUS);
1041
1042         spin_lock(&hugetlb_lock);
1043         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1044         spin_unlock(&hugetlb_lock);
1045
1046         if (!page) {
1047                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1048                 if (!page) {
1049                         hugetlb_put_quota(inode->i_mapping, chg);
1050                         return ERR_PTR(-VM_FAULT_SIGBUS);
1051                 }
1052         }
1053
1054         set_page_private(page, (unsigned long) mapping);
1055
1056         vma_commit_reservation(h, vma, addr);
1057
1058         return page;
1059 }
1060
1061 int __weak alloc_bootmem_huge_page(struct hstate *h)
1062 {
1063         struct huge_bootmem_page *m;
1064         int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1065
1066         while (nr_nodes) {
1067                 void *addr;
1068
1069                 addr = __alloc_bootmem_node_nopanic(
1070                                 NODE_DATA(hstate_next_node_to_alloc(h,
1071                                                 &node_states[N_HIGH_MEMORY])),
1072                                 huge_page_size(h), huge_page_size(h), 0);
1073
1074                 if (addr) {
1075                         /*
1076                          * Use the beginning of the huge page to store the
1077                          * huge_bootmem_page struct (until gather_bootmem
1078                          * puts them into the mem_map).
1079                          */
1080                         m = addr;
1081                         goto found;
1082                 }
1083                 nr_nodes--;
1084         }
1085         return 0;
1086
1087 found:
1088         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1089         /* Put them into a private list first because mem_map is not up yet */
1090         list_add(&m->list, &huge_boot_pages);
1091         m->hstate = h;
1092         return 1;
1093 }
1094
1095 static void prep_compound_huge_page(struct page *page, int order)
1096 {
1097         if (unlikely(order > (MAX_ORDER - 1)))
1098                 prep_compound_gigantic_page(page, order);
1099         else
1100                 prep_compound_page(page, order);
1101 }
1102
1103 /* Put bootmem huge pages into the standard lists after mem_map is up */
1104 static void __init gather_bootmem_prealloc(void)
1105 {
1106         struct huge_bootmem_page *m;
1107
1108         list_for_each_entry(m, &huge_boot_pages, list) {
1109                 struct hstate *h = m->hstate;
1110                 struct page *page;
1111
1112 #ifdef CONFIG_HIGHMEM
1113                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1114                 free_bootmem_late((unsigned long)m,
1115                                   sizeof(struct huge_bootmem_page));
1116 #else
1117                 page = virt_to_page(m);
1118 #endif
1119                 __ClearPageReserved(page);
1120                 WARN_ON(page_count(page) != 1);
1121                 prep_compound_huge_page(page, h->order);
1122                 prep_new_huge_page(h, page, page_to_nid(page));
1123                 /*
1124                  * If we had gigantic hugepages allocated at boot time, we need
1125                  * to restore the 'stolen' pages to totalram_pages in order to
1126                  * fix confusing memory reports from free(1) and another
1127                  * side-effects, like CommitLimit going negative.
1128                  */
1129                 if (h->order > (MAX_ORDER - 1))
1130                         totalram_pages += 1 << h->order;
1131         }
1132 }
1133
1134 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1135 {
1136         unsigned long i;
1137
1138         for (i = 0; i < h->max_huge_pages; ++i) {
1139                 if (h->order >= MAX_ORDER) {
1140                         if (!alloc_bootmem_huge_page(h))
1141                                 break;
1142                 } else if (!alloc_fresh_huge_page(h,
1143                                          &node_states[N_HIGH_MEMORY]))
1144                         break;
1145         }
1146         h->max_huge_pages = i;
1147 }
1148
1149 static void __init hugetlb_init_hstates(void)
1150 {
1151         struct hstate *h;
1152
1153         for_each_hstate(h) {
1154                 /* oversize hugepages were init'ed in early boot */
1155                 if (h->order < MAX_ORDER)
1156                         hugetlb_hstate_alloc_pages(h);
1157         }
1158 }
1159
1160 static char * __init memfmt(char *buf, unsigned long n)
1161 {
1162         if (n >= (1UL << 30))
1163                 sprintf(buf, "%lu GB", n >> 30);
1164         else if (n >= (1UL << 20))
1165                 sprintf(buf, "%lu MB", n >> 20);
1166         else
1167                 sprintf(buf, "%lu KB", n >> 10);
1168         return buf;
1169 }
1170
1171 static void __init report_hugepages(void)
1172 {
1173         struct hstate *h;
1174
1175         for_each_hstate(h) {
1176                 char buf[32];
1177                 printk(KERN_INFO "HugeTLB registered %s page size, "
1178                                  "pre-allocated %ld pages\n",
1179                         memfmt(buf, huge_page_size(h)),
1180                         h->free_huge_pages);
1181         }
1182 }
1183
1184 #ifdef CONFIG_HIGHMEM
1185 static void try_to_free_low(struct hstate *h, unsigned long count,
1186                                                 nodemask_t *nodes_allowed)
1187 {
1188         int i;
1189
1190         if (h->order >= MAX_ORDER)
1191                 return;
1192
1193         for_each_node_mask(i, *nodes_allowed) {
1194                 struct page *page, *next;
1195                 struct list_head *freel = &h->hugepage_freelists[i];
1196                 list_for_each_entry_safe(page, next, freel, lru) {
1197                         if (count >= h->nr_huge_pages)
1198                                 return;
1199                         if (PageHighMem(page))
1200                                 continue;
1201                         list_del(&page->lru);
1202                         update_and_free_page(h, page);
1203                         h->free_huge_pages--;
1204                         h->free_huge_pages_node[page_to_nid(page)]--;
1205                 }
1206         }
1207 }
1208 #else
1209 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1210                                                 nodemask_t *nodes_allowed)
1211 {
1212 }
1213 #endif
1214
1215 /*
1216  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1217  * balanced by operating on them in a round-robin fashion.
1218  * Returns 1 if an adjustment was made.
1219  */
1220 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1221                                 int delta)
1222 {
1223         int start_nid, next_nid;
1224         int ret = 0;
1225
1226         VM_BUG_ON(delta != -1 && delta != 1);
1227
1228         if (delta < 0)
1229                 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1230         else
1231                 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1232         next_nid = start_nid;
1233
1234         do {
1235                 int nid = next_nid;
1236                 if (delta < 0)  {
1237                         /*
1238                          * To shrink on this node, there must be a surplus page
1239                          */
1240                         if (!h->surplus_huge_pages_node[nid]) {
1241                                 next_nid = hstate_next_node_to_alloc(h,
1242                                                                 nodes_allowed);
1243                                 continue;
1244                         }
1245                 }
1246                 if (delta > 0) {
1247                         /*
1248                          * Surplus cannot exceed the total number of pages
1249                          */
1250                         if (h->surplus_huge_pages_node[nid] >=
1251                                                 h->nr_huge_pages_node[nid]) {
1252                                 next_nid = hstate_next_node_to_free(h,
1253                                                                 nodes_allowed);
1254                                 continue;
1255                         }
1256                 }
1257
1258                 h->surplus_huge_pages += delta;
1259                 h->surplus_huge_pages_node[nid] += delta;
1260                 ret = 1;
1261                 break;
1262         } while (next_nid != start_nid);
1263
1264         return ret;
1265 }
1266
1267 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1268 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1269                                                 nodemask_t *nodes_allowed)
1270 {
1271         unsigned long min_count, ret;
1272
1273         if (h->order >= MAX_ORDER)
1274                 return h->max_huge_pages;
1275
1276         /*
1277          * Increase the pool size
1278          * First take pages out of surplus state.  Then make up the
1279          * remaining difference by allocating fresh huge pages.
1280          *
1281          * We might race with alloc_buddy_huge_page() here and be unable
1282          * to convert a surplus huge page to a normal huge page. That is
1283          * not critical, though, it just means the overall size of the
1284          * pool might be one hugepage larger than it needs to be, but
1285          * within all the constraints specified by the sysctls.
1286          */
1287         spin_lock(&hugetlb_lock);
1288         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1289                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1290                         break;
1291         }
1292
1293         while (count > persistent_huge_pages(h)) {
1294                 /*
1295                  * If this allocation races such that we no longer need the
1296                  * page, free_huge_page will handle it by freeing the page
1297                  * and reducing the surplus.
1298                  */
1299                 spin_unlock(&hugetlb_lock);
1300                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1301                 spin_lock(&hugetlb_lock);
1302                 if (!ret)
1303                         goto out;
1304
1305                 /* Bail for signals. Probably ctrl-c from user */
1306                 if (signal_pending(current))
1307                         goto out;
1308         }
1309
1310         /*
1311          * Decrease the pool size
1312          * First return free pages to the buddy allocator (being careful
1313          * to keep enough around to satisfy reservations).  Then place
1314          * pages into surplus state as needed so the pool will shrink
1315          * to the desired size as pages become free.
1316          *
1317          * By placing pages into the surplus state independent of the
1318          * overcommit value, we are allowing the surplus pool size to
1319          * exceed overcommit. There are few sane options here. Since
1320          * alloc_buddy_huge_page() is checking the global counter,
1321          * though, we'll note that we're not allowed to exceed surplus
1322          * and won't grow the pool anywhere else. Not until one of the
1323          * sysctls are changed, or the surplus pages go out of use.
1324          */
1325         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1326         min_count = max(count, min_count);
1327         try_to_free_low(h, min_count, nodes_allowed);
1328         while (min_count < persistent_huge_pages(h)) {
1329                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1330                         break;
1331         }
1332         while (count < persistent_huge_pages(h)) {
1333                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1334                         break;
1335         }
1336 out:
1337         ret = persistent_huge_pages(h);
1338         spin_unlock(&hugetlb_lock);
1339         return ret;
1340 }
1341
1342 #define HSTATE_ATTR_RO(_name) \
1343         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1344
1345 #define HSTATE_ATTR(_name) \
1346         static struct kobj_attribute _name##_attr = \
1347                 __ATTR(_name, 0644, _name##_show, _name##_store)
1348
1349 static struct kobject *hugepages_kobj;
1350 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1351
1352 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1353
1354 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1355 {
1356         int i;
1357
1358         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1359                 if (hstate_kobjs[i] == kobj) {
1360                         if (nidp)
1361                                 *nidp = NUMA_NO_NODE;
1362                         return &hstates[i];
1363                 }
1364
1365         return kobj_to_node_hstate(kobj, nidp);
1366 }
1367
1368 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1369                                         struct kobj_attribute *attr, char *buf)
1370 {
1371         struct hstate *h;
1372         unsigned long nr_huge_pages;
1373         int nid;
1374
1375         h = kobj_to_hstate(kobj, &nid);
1376         if (nid == NUMA_NO_NODE)
1377                 nr_huge_pages = h->nr_huge_pages;
1378         else
1379                 nr_huge_pages = h->nr_huge_pages_node[nid];
1380
1381         return sprintf(buf, "%lu\n", nr_huge_pages);
1382 }
1383
1384 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1385                         struct kobject *kobj, struct kobj_attribute *attr,
1386                         const char *buf, size_t len)
1387 {
1388         int err;
1389         int nid;
1390         unsigned long count;
1391         struct hstate *h;
1392         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1393
1394         err = strict_strtoul(buf, 10, &count);
1395         if (err)
1396                 goto out;
1397
1398         h = kobj_to_hstate(kobj, &nid);
1399         if (h->order >= MAX_ORDER) {
1400                 err = -EINVAL;
1401                 goto out;
1402         }
1403
1404         if (nid == NUMA_NO_NODE) {
1405                 /*
1406                  * global hstate attribute
1407                  */
1408                 if (!(obey_mempolicy &&
1409                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1410                         NODEMASK_FREE(nodes_allowed);
1411                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1412                 }
1413         } else if (nodes_allowed) {
1414                 /*
1415                  * per node hstate attribute: adjust count to global,
1416                  * but restrict alloc/free to the specified node.
1417                  */
1418                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1419                 init_nodemask_of_node(nodes_allowed, nid);
1420         } else
1421                 nodes_allowed = &node_states[N_HIGH_MEMORY];
1422
1423         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1424
1425         if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1426                 NODEMASK_FREE(nodes_allowed);
1427
1428         return len;
1429 out:
1430         NODEMASK_FREE(nodes_allowed);
1431         return err;
1432 }
1433
1434 static ssize_t nr_hugepages_show(struct kobject *kobj,
1435                                        struct kobj_attribute *attr, char *buf)
1436 {
1437         return nr_hugepages_show_common(kobj, attr, buf);
1438 }
1439
1440 static ssize_t nr_hugepages_store(struct kobject *kobj,
1441                struct kobj_attribute *attr, const char *buf, size_t len)
1442 {
1443         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1444 }
1445 HSTATE_ATTR(nr_hugepages);
1446
1447 #ifdef CONFIG_NUMA
1448
1449 /*
1450  * hstate attribute for optionally mempolicy-based constraint on persistent
1451  * huge page alloc/free.
1452  */
1453 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1454                                        struct kobj_attribute *attr, char *buf)
1455 {
1456         return nr_hugepages_show_common(kobj, attr, buf);
1457 }
1458
1459 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1460                struct kobj_attribute *attr, const char *buf, size_t len)
1461 {
1462         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1463 }
1464 HSTATE_ATTR(nr_hugepages_mempolicy);
1465 #endif
1466
1467
1468 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1469                                         struct kobj_attribute *attr, char *buf)
1470 {
1471         struct hstate *h = kobj_to_hstate(kobj, NULL);
1472         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1473 }
1474
1475 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1476                 struct kobj_attribute *attr, const char *buf, size_t count)
1477 {
1478         int err;
1479         unsigned long input;
1480         struct hstate *h = kobj_to_hstate(kobj, NULL);
1481
1482         if (h->order >= MAX_ORDER)
1483                 return -EINVAL;
1484
1485         err = strict_strtoul(buf, 10, &input);
1486         if (err)
1487                 return err;
1488
1489         spin_lock(&hugetlb_lock);
1490         h->nr_overcommit_huge_pages = input;
1491         spin_unlock(&hugetlb_lock);
1492
1493         return count;
1494 }
1495 HSTATE_ATTR(nr_overcommit_hugepages);
1496
1497 static ssize_t free_hugepages_show(struct kobject *kobj,
1498                                         struct kobj_attribute *attr, char *buf)
1499 {
1500         struct hstate *h;
1501         unsigned long free_huge_pages;
1502         int nid;
1503
1504         h = kobj_to_hstate(kobj, &nid);
1505         if (nid == NUMA_NO_NODE)
1506                 free_huge_pages = h->free_huge_pages;
1507         else
1508                 free_huge_pages = h->free_huge_pages_node[nid];
1509
1510         return sprintf(buf, "%lu\n", free_huge_pages);
1511 }
1512 HSTATE_ATTR_RO(free_hugepages);
1513
1514 static ssize_t resv_hugepages_show(struct kobject *kobj,
1515                                         struct kobj_attribute *attr, char *buf)
1516 {
1517         struct hstate *h = kobj_to_hstate(kobj, NULL);
1518         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1519 }
1520 HSTATE_ATTR_RO(resv_hugepages);
1521
1522 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1523                                         struct kobj_attribute *attr, char *buf)
1524 {
1525         struct hstate *h;
1526         unsigned long surplus_huge_pages;
1527         int nid;
1528
1529         h = kobj_to_hstate(kobj, &nid);
1530         if (nid == NUMA_NO_NODE)
1531                 surplus_huge_pages = h->surplus_huge_pages;
1532         else
1533                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1534
1535         return sprintf(buf, "%lu\n", surplus_huge_pages);
1536 }
1537 HSTATE_ATTR_RO(surplus_hugepages);
1538
1539 static struct attribute *hstate_attrs[] = {
1540         &nr_hugepages_attr.attr,
1541         &nr_overcommit_hugepages_attr.attr,
1542         &free_hugepages_attr.attr,
1543         &resv_hugepages_attr.attr,
1544         &surplus_hugepages_attr.attr,
1545 #ifdef CONFIG_NUMA
1546         &nr_hugepages_mempolicy_attr.attr,
1547 #endif
1548         NULL,
1549 };
1550
1551 static struct attribute_group hstate_attr_group = {
1552         .attrs = hstate_attrs,
1553 };
1554
1555 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1556                                     struct kobject **hstate_kobjs,
1557                                     struct attribute_group *hstate_attr_group)
1558 {
1559         int retval;
1560         int hi = h - hstates;
1561
1562         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1563         if (!hstate_kobjs[hi])
1564                 return -ENOMEM;
1565
1566         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1567         if (retval)
1568                 kobject_put(hstate_kobjs[hi]);
1569
1570         return retval;
1571 }
1572
1573 static void __init hugetlb_sysfs_init(void)
1574 {
1575         struct hstate *h;
1576         int err;
1577
1578         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1579         if (!hugepages_kobj)
1580                 return;
1581
1582         for_each_hstate(h) {
1583                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1584                                          hstate_kobjs, &hstate_attr_group);
1585                 if (err)
1586                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1587                                                                 h->name);
1588         }
1589 }
1590
1591 #ifdef CONFIG_NUMA
1592
1593 /*
1594  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1595  * with node devices in node_devices[] using a parallel array.  The array
1596  * index of a node device or _hstate == node id.
1597  * This is here to avoid any static dependency of the node device driver, in
1598  * the base kernel, on the hugetlb module.
1599  */
1600 struct node_hstate {
1601         struct kobject          *hugepages_kobj;
1602         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1603 };
1604 struct node_hstate node_hstates[MAX_NUMNODES];
1605
1606 /*
1607  * A subset of global hstate attributes for node devices
1608  */
1609 static struct attribute *per_node_hstate_attrs[] = {
1610         &nr_hugepages_attr.attr,
1611         &free_hugepages_attr.attr,
1612         &surplus_hugepages_attr.attr,
1613         NULL,
1614 };
1615
1616 static struct attribute_group per_node_hstate_attr_group = {
1617         .attrs = per_node_hstate_attrs,
1618 };
1619
1620 /*
1621  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1622  * Returns node id via non-NULL nidp.
1623  */
1624 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1625 {
1626         int nid;
1627
1628         for (nid = 0; nid < nr_node_ids; nid++) {
1629                 struct node_hstate *nhs = &node_hstates[nid];
1630                 int i;
1631                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1632                         if (nhs->hstate_kobjs[i] == kobj) {
1633                                 if (nidp)
1634                                         *nidp = nid;
1635                                 return &hstates[i];
1636                         }
1637         }
1638
1639         BUG();
1640         return NULL;
1641 }
1642
1643 /*
1644  * Unregister hstate attributes from a single node device.
1645  * No-op if no hstate attributes attached.
1646  */
1647 void hugetlb_unregister_node(struct node *node)
1648 {
1649         struct hstate *h;
1650         struct node_hstate *nhs = &node_hstates[node->dev.id];
1651
1652         if (!nhs->hugepages_kobj)
1653                 return;         /* no hstate attributes */
1654
1655         for_each_hstate(h)
1656                 if (nhs->hstate_kobjs[h - hstates]) {
1657                         kobject_put(nhs->hstate_kobjs[h - hstates]);
1658                         nhs->hstate_kobjs[h - hstates] = NULL;
1659                 }
1660
1661         kobject_put(nhs->hugepages_kobj);
1662         nhs->hugepages_kobj = NULL;
1663 }
1664
1665 /*
1666  * hugetlb module exit:  unregister hstate attributes from node devices
1667  * that have them.
1668  */
1669 static void hugetlb_unregister_all_nodes(void)
1670 {
1671         int nid;
1672
1673         /*
1674          * disable node device registrations.
1675          */
1676         register_hugetlbfs_with_node(NULL, NULL);
1677
1678         /*
1679          * remove hstate attributes from any nodes that have them.
1680          */
1681         for (nid = 0; nid < nr_node_ids; nid++)
1682                 hugetlb_unregister_node(&node_devices[nid]);
1683 }
1684
1685 /*
1686  * Register hstate attributes for a single node device.
1687  * No-op if attributes already registered.
1688  */
1689 void hugetlb_register_node(struct node *node)
1690 {
1691         struct hstate *h;
1692         struct node_hstate *nhs = &node_hstates[node->dev.id];
1693         int err;
1694
1695         if (nhs->hugepages_kobj)
1696                 return;         /* already allocated */
1697
1698         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1699                                                         &node->dev.kobj);
1700         if (!nhs->hugepages_kobj)
1701                 return;
1702
1703         for_each_hstate(h) {
1704                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1705                                                 nhs->hstate_kobjs,
1706                                                 &per_node_hstate_attr_group);
1707                 if (err) {
1708                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1709                                         " for node %d\n",
1710                                                 h->name, node->dev.id);
1711                         hugetlb_unregister_node(node);
1712                         break;
1713                 }
1714         }
1715 }
1716
1717 /*
1718  * hugetlb init time:  register hstate attributes for all registered node
1719  * devices of nodes that have memory.  All on-line nodes should have
1720  * registered their associated device by this time.
1721  */
1722 static void hugetlb_register_all_nodes(void)
1723 {
1724         int nid;
1725
1726         for_each_node_state(nid, N_HIGH_MEMORY) {
1727                 struct node *node = &node_devices[nid];
1728                 if (node->dev.id == nid)
1729                         hugetlb_register_node(node);
1730         }
1731
1732         /*
1733          * Let the node device driver know we're here so it can
1734          * [un]register hstate attributes on node hotplug.
1735          */
1736         register_hugetlbfs_with_node(hugetlb_register_node,
1737                                      hugetlb_unregister_node);
1738 }
1739 #else   /* !CONFIG_NUMA */
1740
1741 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1742 {
1743         BUG();
1744         if (nidp)
1745                 *nidp = -1;
1746         return NULL;
1747 }
1748
1749 static void hugetlb_unregister_all_nodes(void) { }
1750
1751 static void hugetlb_register_all_nodes(void) { }
1752
1753 #endif
1754
1755 static void __exit hugetlb_exit(void)
1756 {
1757         struct hstate *h;
1758
1759         hugetlb_unregister_all_nodes();
1760
1761         for_each_hstate(h) {
1762                 kobject_put(hstate_kobjs[h - hstates]);
1763         }
1764
1765         kobject_put(hugepages_kobj);
1766 }
1767 module_exit(hugetlb_exit);
1768
1769 static int __init hugetlb_init(void)
1770 {
1771         /* Some platform decide whether they support huge pages at boot
1772          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1773          * there is no such support
1774          */
1775         if (HPAGE_SHIFT == 0)
1776                 return 0;
1777
1778         if (!size_to_hstate(default_hstate_size)) {
1779                 default_hstate_size = HPAGE_SIZE;
1780                 if (!size_to_hstate(default_hstate_size))
1781                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1782         }
1783         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1784         if (default_hstate_max_huge_pages)
1785                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1786
1787         hugetlb_init_hstates();
1788
1789         gather_bootmem_prealloc();
1790
1791         report_hugepages();
1792
1793         hugetlb_sysfs_init();
1794
1795         hugetlb_register_all_nodes();
1796
1797         return 0;
1798 }
1799 module_init(hugetlb_init);
1800
1801 /* Should be called on processing a hugepagesz=... option */
1802 void __init hugetlb_add_hstate(unsigned order)
1803 {
1804         struct hstate *h;
1805         unsigned long i;
1806
1807         if (size_to_hstate(PAGE_SIZE << order)) {
1808                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1809                 return;
1810         }
1811         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1812         BUG_ON(order == 0);
1813         h = &hstates[max_hstate++];
1814         h->order = order;
1815         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1816         h->nr_huge_pages = 0;
1817         h->free_huge_pages = 0;
1818         for (i = 0; i < MAX_NUMNODES; ++i)
1819                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1820         h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1821         h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1822         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1823                                         huge_page_size(h)/1024);
1824
1825         parsed_hstate = h;
1826 }
1827
1828 static int __init hugetlb_nrpages_setup(char *s)
1829 {
1830         unsigned long *mhp;
1831         static unsigned long *last_mhp;
1832
1833         /*
1834          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1835          * so this hugepages= parameter goes to the "default hstate".
1836          */
1837         if (!max_hstate)
1838                 mhp = &default_hstate_max_huge_pages;
1839         else
1840                 mhp = &parsed_hstate->max_huge_pages;
1841
1842         if (mhp == last_mhp) {
1843                 printk(KERN_WARNING "hugepages= specified twice without "
1844                         "interleaving hugepagesz=, ignoring\n");
1845                 return 1;
1846         }
1847
1848         if (sscanf(s, "%lu", mhp) <= 0)
1849                 *mhp = 0;
1850
1851         /*
1852          * Global state is always initialized later in hugetlb_init.
1853          * But we need to allocate >= MAX_ORDER hstates here early to still
1854          * use the bootmem allocator.
1855          */
1856         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1857                 hugetlb_hstate_alloc_pages(parsed_hstate);
1858
1859         last_mhp = mhp;
1860
1861         return 1;
1862 }
1863 __setup("hugepages=", hugetlb_nrpages_setup);
1864
1865 static int __init hugetlb_default_setup(char *s)
1866 {
1867         default_hstate_size = memparse(s, &s);
1868         return 1;
1869 }
1870 __setup("default_hugepagesz=", hugetlb_default_setup);
1871
1872 static unsigned int cpuset_mems_nr(unsigned int *array)
1873 {
1874         int node;
1875         unsigned int nr = 0;
1876
1877         for_each_node_mask(node, cpuset_current_mems_allowed)
1878                 nr += array[node];
1879
1880         return nr;
1881 }
1882
1883 #ifdef CONFIG_SYSCTL
1884 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1885                          struct ctl_table *table, int write,
1886                          void __user *buffer, size_t *length, loff_t *ppos)
1887 {
1888         struct hstate *h = &default_hstate;
1889         unsigned long tmp;
1890         int ret;
1891
1892         tmp = h->max_huge_pages;
1893
1894         if (write && h->order >= MAX_ORDER)
1895                 return -EINVAL;
1896
1897         table->data = &tmp;
1898         table->maxlen = sizeof(unsigned long);
1899         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1900         if (ret)
1901                 goto out;
1902
1903         if (write) {
1904                 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1905                                                 GFP_KERNEL | __GFP_NORETRY);
1906                 if (!(obey_mempolicy &&
1907                                init_nodemask_of_mempolicy(nodes_allowed))) {
1908                         NODEMASK_FREE(nodes_allowed);
1909                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1910                 }
1911                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1912
1913                 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1914                         NODEMASK_FREE(nodes_allowed);
1915         }
1916 out:
1917         return ret;
1918 }
1919
1920 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1921                           void __user *buffer, size_t *length, loff_t *ppos)
1922 {
1923
1924         return hugetlb_sysctl_handler_common(false, table, write,
1925                                                         buffer, length, ppos);
1926 }
1927
1928 #ifdef CONFIG_NUMA
1929 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1930                           void __user *buffer, size_t *length, loff_t *ppos)
1931 {
1932         return hugetlb_sysctl_handler_common(true, table, write,
1933                                                         buffer, length, ppos);
1934 }
1935 #endif /* CONFIG_NUMA */
1936
1937 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1938                         void __user *buffer,
1939                         size_t *length, loff_t *ppos)
1940 {
1941         proc_dointvec(table, write, buffer, length, ppos);
1942         if (hugepages_treat_as_movable)
1943                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1944         else
1945                 htlb_alloc_mask = GFP_HIGHUSER;
1946         return 0;
1947 }
1948
1949 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1950                         void __user *buffer,
1951                         size_t *length, loff_t *ppos)
1952 {
1953         struct hstate *h = &default_hstate;
1954         unsigned long tmp;
1955         int ret;
1956
1957         tmp = h->nr_overcommit_huge_pages;
1958
1959         if (write && h->order >= MAX_ORDER)
1960                 return -EINVAL;
1961
1962         table->data = &tmp;
1963         table->maxlen = sizeof(unsigned long);
1964         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1965         if (ret)
1966                 goto out;
1967
1968         if (write) {
1969                 spin_lock(&hugetlb_lock);
1970                 h->nr_overcommit_huge_pages = tmp;
1971                 spin_unlock(&hugetlb_lock);
1972         }
1973 out:
1974         return ret;
1975 }
1976
1977 #endif /* CONFIG_SYSCTL */
1978
1979 void hugetlb_report_meminfo(struct seq_file *m)
1980 {
1981         struct hstate *h = &default_hstate;
1982         seq_printf(m,
1983                         "HugePages_Total:   %5lu\n"
1984                         "HugePages_Free:    %5lu\n"
1985                         "HugePages_Rsvd:    %5lu\n"
1986                         "HugePages_Surp:    %5lu\n"
1987                         "Hugepagesize:   %8lu kB\n",
1988                         h->nr_huge_pages,
1989                         h->free_huge_pages,
1990                         h->resv_huge_pages,
1991                         h->surplus_huge_pages,
1992                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1993 }
1994
1995 int hugetlb_report_node_meminfo(int nid, char *buf)
1996 {
1997         struct hstate *h = &default_hstate;
1998         return sprintf(buf,
1999                 "Node %d HugePages_Total: %5u\n"
2000                 "Node %d HugePages_Free:  %5u\n"
2001                 "Node %d HugePages_Surp:  %5u\n",
2002                 nid, h->nr_huge_pages_node[nid],
2003                 nid, h->free_huge_pages_node[nid],
2004                 nid, h->surplus_huge_pages_node[nid]);
2005 }
2006
2007 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2008 unsigned long hugetlb_total_pages(void)
2009 {
2010         struct hstate *h = &default_hstate;
2011         return h->nr_huge_pages * pages_per_huge_page(h);
2012 }
2013
2014 static int hugetlb_acct_memory(struct hstate *h, long delta)
2015 {
2016         int ret = -ENOMEM;
2017
2018         spin_lock(&hugetlb_lock);
2019         /*
2020          * When cpuset is configured, it breaks the strict hugetlb page
2021          * reservation as the accounting is done on a global variable. Such
2022          * reservation is completely rubbish in the presence of cpuset because
2023          * the reservation is not checked against page availability for the
2024          * current cpuset. Application can still potentially OOM'ed by kernel
2025          * with lack of free htlb page in cpuset that the task is in.
2026          * Attempt to enforce strict accounting with cpuset is almost
2027          * impossible (or too ugly) because cpuset is too fluid that
2028          * task or memory node can be dynamically moved between cpusets.
2029          *
2030          * The change of semantics for shared hugetlb mapping with cpuset is
2031          * undesirable. However, in order to preserve some of the semantics,
2032          * we fall back to check against current free page availability as
2033          * a best attempt and hopefully to minimize the impact of changing
2034          * semantics that cpuset has.
2035          */
2036         if (delta > 0) {
2037                 if (gather_surplus_pages(h, delta) < 0)
2038                         goto out;
2039
2040                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2041                         return_unused_surplus_pages(h, delta);
2042                         goto out;
2043                 }
2044         }
2045
2046         ret = 0;
2047         if (delta < 0)
2048                 return_unused_surplus_pages(h, (unsigned long) -delta);
2049
2050 out:
2051         spin_unlock(&hugetlb_lock);
2052         return ret;
2053 }
2054
2055 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2056 {
2057         struct resv_map *reservations = vma_resv_map(vma);
2058
2059         /*
2060          * This new VMA should share its siblings reservation map if present.
2061          * The VMA will only ever have a valid reservation map pointer where
2062          * it is being copied for another still existing VMA.  As that VMA
2063          * has a reference to the reservation map it cannot disappear until
2064          * after this open call completes.  It is therefore safe to take a
2065          * new reference here without additional locking.
2066          */
2067         if (reservations)
2068                 kref_get(&reservations->refs);
2069 }
2070
2071 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2072 {
2073         struct hstate *h = hstate_vma(vma);
2074         struct resv_map *reservations = vma_resv_map(vma);
2075         unsigned long reserve;
2076         unsigned long start;
2077         unsigned long end;
2078
2079         if (reservations) {
2080                 start = vma_hugecache_offset(h, vma, vma->vm_start);
2081                 end = vma_hugecache_offset(h, vma, vma->vm_end);
2082
2083                 reserve = (end - start) -
2084                         region_count(&reservations->regions, start, end);
2085
2086                 kref_put(&reservations->refs, resv_map_release);
2087
2088                 if (reserve) {
2089                         hugetlb_acct_memory(h, -reserve);
2090                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2091                 }
2092         }
2093 }
2094
2095 /*
2096  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2097  * handle_mm_fault() to try to instantiate regular-sized pages in the
2098  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2099  * this far.
2100  */
2101 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2102 {
2103         BUG();
2104         return 0;
2105 }
2106
2107 const struct vm_operations_struct hugetlb_vm_ops = {
2108         .fault = hugetlb_vm_op_fault,
2109         .open = hugetlb_vm_op_open,
2110         .close = hugetlb_vm_op_close,
2111 };
2112
2113 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2114                                 int writable)
2115 {
2116         pte_t entry;
2117
2118         if (writable) {
2119                 entry =
2120                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2121         } else {
2122                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2123         }
2124         entry = pte_mkyoung(entry);
2125         entry = pte_mkhuge(entry);
2126
2127         return entry;
2128 }
2129
2130 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2131                                    unsigned long address, pte_t *ptep)
2132 {
2133         pte_t entry;
2134
2135         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2136         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2137                 update_mmu_cache(vma, address, ptep);
2138 }
2139
2140
2141 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2142                             struct vm_area_struct *vma)
2143 {
2144         pte_t *src_pte, *dst_pte, entry;
2145         struct page *ptepage;
2146         unsigned long addr;
2147         int cow;
2148         struct hstate *h = hstate_vma(vma);
2149         unsigned long sz = huge_page_size(h);
2150
2151         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2152
2153         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2154                 src_pte = huge_pte_offset(src, addr);
2155                 if (!src_pte)
2156                         continue;
2157                 dst_pte = huge_pte_alloc(dst, addr, sz);
2158                 if (!dst_pte)
2159                         goto nomem;
2160
2161                 /* If the pagetables are shared don't copy or take references */
2162                 if (dst_pte == src_pte)
2163                         continue;
2164
2165                 spin_lock(&dst->page_table_lock);
2166                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2167                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2168                         if (cow)
2169                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2170                         entry = huge_ptep_get(src_pte);
2171                         ptepage = pte_page(entry);
2172                         get_page(ptepage);
2173                         page_dup_rmap(ptepage);
2174                         set_huge_pte_at(dst, addr, dst_pte, entry);
2175                 }
2176                 spin_unlock(&src->page_table_lock);
2177                 spin_unlock(&dst->page_table_lock);
2178         }
2179         return 0;
2180
2181 nomem:
2182         return -ENOMEM;
2183 }
2184
2185 static int is_hugetlb_entry_migration(pte_t pte)
2186 {
2187         swp_entry_t swp;
2188
2189         if (huge_pte_none(pte) || pte_present(pte))
2190                 return 0;
2191         swp = pte_to_swp_entry(pte);
2192         if (non_swap_entry(swp) && is_migration_entry(swp))
2193                 return 1;
2194         else
2195                 return 0;
2196 }
2197
2198 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2199 {
2200         swp_entry_t swp;
2201
2202         if (huge_pte_none(pte) || pte_present(pte))
2203                 return 0;
2204         swp = pte_to_swp_entry(pte);
2205         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2206                 return 1;
2207         else
2208                 return 0;
2209 }
2210
2211 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2212                             unsigned long end, struct page *ref_page)
2213 {
2214         struct mm_struct *mm = vma->vm_mm;
2215         unsigned long address;
2216         pte_t *ptep;
2217         pte_t pte;
2218         struct page *page;
2219         struct page *tmp;
2220         struct hstate *h = hstate_vma(vma);
2221         unsigned long sz = huge_page_size(h);
2222
2223         /*
2224          * A page gathering list, protected by per file i_mmap_mutex. The
2225          * lock is used to avoid list corruption from multiple unmapping
2226          * of the same page since we are using page->lru.
2227          */
2228         LIST_HEAD(page_list);
2229
2230         WARN_ON(!is_vm_hugetlb_page(vma));
2231         BUG_ON(start & ~huge_page_mask(h));
2232         BUG_ON(end & ~huge_page_mask(h));
2233
2234         mmu_notifier_invalidate_range_start(mm, start, end);
2235         spin_lock(&mm->page_table_lock);
2236         for (address = start; address < end; address += sz) {
2237                 ptep = huge_pte_offset(mm, address);
2238                 if (!ptep)
2239                         continue;
2240
2241                 if (huge_pmd_unshare(mm, &address, ptep))
2242                         continue;
2243
2244                 /*
2245                  * If a reference page is supplied, it is because a specific
2246                  * page is being unmapped, not a range. Ensure the page we
2247                  * are about to unmap is the actual page of interest.
2248                  */
2249                 if (ref_page) {
2250                         pte = huge_ptep_get(ptep);
2251                         if (huge_pte_none(pte))
2252                                 continue;
2253                         page = pte_page(pte);
2254                         if (page != ref_page)
2255                                 continue;
2256
2257                         /*
2258                          * Mark the VMA as having unmapped its page so that
2259                          * future faults in this VMA will fail rather than
2260                          * looking like data was lost
2261                          */
2262                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2263                 }
2264
2265                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2266                 if (huge_pte_none(pte))
2267                         continue;
2268
2269                 /*
2270                  * HWPoisoned hugepage is already unmapped and dropped reference
2271                  */
2272                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2273                         continue;
2274
2275                 page = pte_page(pte);
2276                 if (pte_dirty(pte))
2277                         set_page_dirty(page);
2278                 list_add(&page->lru, &page_list);
2279         }
2280         spin_unlock(&mm->page_table_lock);
2281         flush_tlb_range(vma, start, end);
2282         mmu_notifier_invalidate_range_end(mm, start, end);
2283         list_for_each_entry_safe(page, tmp, &page_list, lru) {
2284                 page_remove_rmap(page);
2285                 list_del(&page->lru);
2286                 put_page(page);
2287         }
2288 }
2289
2290 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2291                           unsigned long end, struct page *ref_page)
2292 {
2293         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2294         __unmap_hugepage_range(vma, start, end, ref_page);
2295         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2296 }
2297
2298 /*
2299  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2300  * mappping it owns the reserve page for. The intention is to unmap the page
2301  * from other VMAs and let the children be SIGKILLed if they are faulting the
2302  * same region.
2303  */
2304 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2305                                 struct page *page, unsigned long address)
2306 {
2307         struct hstate *h = hstate_vma(vma);
2308         struct vm_area_struct *iter_vma;
2309         struct address_space *mapping;
2310         struct prio_tree_iter iter;
2311         pgoff_t pgoff;
2312
2313         /*
2314          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2315          * from page cache lookup which is in HPAGE_SIZE units.
2316          */
2317         address = address & huge_page_mask(h);
2318         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2319                 + (vma->vm_pgoff >> PAGE_SHIFT);
2320         mapping = (struct address_space *)page_private(page);
2321
2322         /*
2323          * Take the mapping lock for the duration of the table walk. As
2324          * this mapping should be shared between all the VMAs,
2325          * __unmap_hugepage_range() is called as the lock is already held
2326          */
2327         mutex_lock(&mapping->i_mmap_mutex);
2328         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2329                 /* Do not unmap the current VMA */
2330                 if (iter_vma == vma)
2331                         continue;
2332
2333                 /*
2334                  * Unmap the page from other VMAs without their own reserves.
2335                  * They get marked to be SIGKILLed if they fault in these
2336                  * areas. This is because a future no-page fault on this VMA
2337                  * could insert a zeroed page instead of the data existing
2338                  * from the time of fork. This would look like data corruption
2339                  */
2340                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2341                         __unmap_hugepage_range(iter_vma,
2342                                 address, address + huge_page_size(h),
2343                                 page);
2344         }
2345         mutex_unlock(&mapping->i_mmap_mutex);
2346
2347         return 1;
2348 }
2349
2350 /*
2351  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2352  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2353  * cannot race with other handlers or page migration.
2354  * Keep the pte_same checks anyway to make transition from the mutex easier.
2355  */
2356 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2357                         unsigned long address, pte_t *ptep, pte_t pte,
2358                         struct page *pagecache_page)
2359 {
2360         struct hstate *h = hstate_vma(vma);
2361         struct page *old_page, *new_page;
2362         int avoidcopy;
2363         int outside_reserve = 0;
2364
2365         old_page = pte_page(pte);
2366
2367 retry_avoidcopy:
2368         /* If no-one else is actually using this page, avoid the copy
2369          * and just make the page writable */
2370         avoidcopy = (page_mapcount(old_page) == 1);
2371         if (avoidcopy) {
2372                 if (PageAnon(old_page))
2373                         page_move_anon_rmap(old_page, vma, address);
2374                 set_huge_ptep_writable(vma, address, ptep);
2375                 return 0;
2376         }
2377
2378         /*
2379          * If the process that created a MAP_PRIVATE mapping is about to
2380          * perform a COW due to a shared page count, attempt to satisfy
2381          * the allocation without using the existing reserves. The pagecache
2382          * page is used to determine if the reserve at this address was
2383          * consumed or not. If reserves were used, a partial faulted mapping
2384          * at the time of fork() could consume its reserves on COW instead
2385          * of the full address range.
2386          */
2387         if (!(vma->vm_flags & VM_MAYSHARE) &&
2388                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2389                         old_page != pagecache_page)
2390                 outside_reserve = 1;
2391
2392         page_cache_get(old_page);
2393
2394         /* Drop page_table_lock as buddy allocator may be called */
2395         spin_unlock(&mm->page_table_lock);
2396         new_page = alloc_huge_page(vma, address, outside_reserve);
2397
2398         if (IS_ERR(new_page)) {
2399                 page_cache_release(old_page);
2400
2401                 /*
2402                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2403                  * it is due to references held by a child and an insufficient
2404                  * huge page pool. To guarantee the original mappers
2405                  * reliability, unmap the page from child processes. The child
2406                  * may get SIGKILLed if it later faults.
2407                  */
2408                 if (outside_reserve) {
2409                         BUG_ON(huge_pte_none(pte));
2410                         if (unmap_ref_private(mm, vma, old_page, address)) {
2411                                 BUG_ON(page_count(old_page) != 1);
2412                                 BUG_ON(huge_pte_none(pte));
2413                                 spin_lock(&mm->page_table_lock);
2414                                 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2415                                 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2416                                         goto retry_avoidcopy;
2417                                 /*
2418                                  * race occurs while re-acquiring page_table_lock, and
2419                                  * our job is done.
2420                                  */
2421                                 return 0;
2422                         }
2423                         WARN_ON_ONCE(1);
2424                 }
2425
2426                 /* Caller expects lock to be held */
2427                 spin_lock(&mm->page_table_lock);
2428                 return -PTR_ERR(new_page);
2429         }
2430
2431         /*
2432          * When the original hugepage is shared one, it does not have
2433          * anon_vma prepared.
2434          */
2435         if (unlikely(anon_vma_prepare(vma))) {
2436                 page_cache_release(new_page);
2437                 page_cache_release(old_page);
2438                 /* Caller expects lock to be held */
2439                 spin_lock(&mm->page_table_lock);
2440                 return VM_FAULT_OOM;
2441         }
2442
2443         copy_user_huge_page(new_page, old_page, address, vma,
2444                             pages_per_huge_page(h));
2445         __SetPageUptodate(new_page);
2446
2447         /*
2448          * Retake the page_table_lock to check for racing updates
2449          * before the page tables are altered
2450          */
2451         spin_lock(&mm->page_table_lock);
2452         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2453         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2454                 /* Break COW */
2455                 mmu_notifier_invalidate_range_start(mm,
2456                         address & huge_page_mask(h),
2457                         (address & huge_page_mask(h)) + huge_page_size(h));
2458                 huge_ptep_clear_flush(vma, address, ptep);
2459                 set_huge_pte_at(mm, address, ptep,
2460                                 make_huge_pte(vma, new_page, 1));
2461                 page_remove_rmap(old_page);
2462                 hugepage_add_new_anon_rmap(new_page, vma, address);
2463                 /* Make the old page be freed below */
2464                 new_page = old_page;
2465                 mmu_notifier_invalidate_range_end(mm,
2466                         address & huge_page_mask(h),
2467                         (address & huge_page_mask(h)) + huge_page_size(h));
2468         }
2469         page_cache_release(new_page);
2470         page_cache_release(old_page);
2471         return 0;
2472 }
2473
2474 /* Return the pagecache page at a given address within a VMA */
2475 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2476                         struct vm_area_struct *vma, unsigned long address)
2477 {
2478         struct address_space *mapping;
2479         pgoff_t idx;
2480
2481         mapping = vma->vm_file->f_mapping;
2482         idx = vma_hugecache_offset(h, vma, address);
2483
2484         return find_lock_page(mapping, idx);
2485 }
2486
2487 /*
2488  * Return whether there is a pagecache page to back given address within VMA.
2489  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2490  */
2491 static bool hugetlbfs_pagecache_present(struct hstate *h,
2492                         struct vm_area_struct *vma, unsigned long address)
2493 {
2494         struct address_space *mapping;
2495         pgoff_t idx;
2496         struct page *page;
2497
2498         mapping = vma->vm_file->f_mapping;
2499         idx = vma_hugecache_offset(h, vma, address);
2500
2501         page = find_get_page(mapping, idx);
2502         if (page)
2503                 put_page(page);
2504         return page != NULL;
2505 }
2506
2507 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2508                         unsigned long address, pte_t *ptep, unsigned int flags)
2509 {
2510         struct hstate *h = hstate_vma(vma);
2511         int ret = VM_FAULT_SIGBUS;
2512         pgoff_t idx;
2513         unsigned long size;
2514         struct page *page;
2515         struct address_space *mapping;
2516         pte_t new_pte;
2517
2518         /*
2519          * Currently, we are forced to kill the process in the event the
2520          * original mapper has unmapped pages from the child due to a failed
2521          * COW. Warn that such a situation has occurred as it may not be obvious
2522          */
2523         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2524                 printk(KERN_WARNING
2525                         "PID %d killed due to inadequate hugepage pool\n",
2526                         current->pid);
2527                 return ret;
2528         }
2529
2530         mapping = vma->vm_file->f_mapping;
2531         idx = vma_hugecache_offset(h, vma, address);
2532
2533         /*
2534          * Use page lock to guard against racing truncation
2535          * before we get page_table_lock.
2536          */
2537 retry:
2538         page = find_lock_page(mapping, idx);
2539         if (!page) {
2540                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2541                 if (idx >= size)
2542                         goto out;
2543                 page = alloc_huge_page(vma, address, 0);
2544                 if (IS_ERR(page)) {
2545                         ret = -PTR_ERR(page);
2546                         goto out;
2547                 }
2548                 clear_huge_page(page, address, pages_per_huge_page(h));
2549                 __SetPageUptodate(page);
2550
2551                 if (vma->vm_flags & VM_MAYSHARE) {
2552                         int err;
2553                         struct inode *inode = mapping->host;
2554
2555                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2556                         if (err) {
2557                                 put_page(page);
2558                                 if (err == -EEXIST)
2559                                         goto retry;
2560                                 goto out;
2561                         }
2562
2563                         spin_lock(&inode->i_lock);
2564                         inode->i_blocks += blocks_per_huge_page(h);
2565                         spin_unlock(&inode->i_lock);
2566                         page_dup_rmap(page);
2567                 } else {
2568                         lock_page(page);
2569                         if (unlikely(anon_vma_prepare(vma))) {
2570                                 ret = VM_FAULT_OOM;
2571                                 goto backout_unlocked;
2572                         }
2573                         hugepage_add_new_anon_rmap(page, vma, address);
2574                 }
2575         } else {
2576                 /*
2577                  * If memory error occurs between mmap() and fault, some process
2578                  * don't have hwpoisoned swap entry for errored virtual address.
2579                  * So we need to block hugepage fault by PG_hwpoison bit check.
2580                  */
2581                 if (unlikely(PageHWPoison(page))) {
2582                         ret = VM_FAULT_HWPOISON |
2583                               VM_FAULT_SET_HINDEX(h - hstates);
2584                         goto backout_unlocked;
2585                 }
2586                 page_dup_rmap(page);
2587         }
2588
2589         /*
2590          * If we are going to COW a private mapping later, we examine the
2591          * pending reservations for this page now. This will ensure that
2592          * any allocations necessary to record that reservation occur outside
2593          * the spinlock.
2594          */
2595         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2596                 if (vma_needs_reservation(h, vma, address) < 0) {
2597                         ret = VM_FAULT_OOM;
2598                         goto backout_unlocked;
2599                 }
2600
2601         spin_lock(&mm->page_table_lock);
2602         size = i_size_read(mapping->host) >> huge_page_shift(h);
2603         if (idx >= size)
2604                 goto backout;
2605
2606         ret = 0;
2607         if (!huge_pte_none(huge_ptep_get(ptep)))
2608                 goto backout;
2609
2610         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2611                                 && (vma->vm_flags & VM_SHARED)));
2612         set_huge_pte_at(mm, address, ptep, new_pte);
2613
2614         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2615                 /* Optimization, do the COW without a second fault */
2616                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2617         }
2618
2619         spin_unlock(&mm->page_table_lock);
2620         unlock_page(page);
2621 out:
2622         return ret;
2623
2624 backout:
2625         spin_unlock(&mm->page_table_lock);
2626 backout_unlocked:
2627         unlock_page(page);
2628         put_page(page);
2629         goto out;
2630 }
2631
2632 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2633                         unsigned long address, unsigned int flags)
2634 {
2635         pte_t *ptep;
2636         pte_t entry;
2637         int ret;
2638         struct page *page = NULL;
2639         struct page *pagecache_page = NULL;
2640         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2641         struct hstate *h = hstate_vma(vma);
2642
2643         address &= huge_page_mask(h);
2644
2645         ptep = huge_pte_offset(mm, address);
2646         if (ptep) {
2647                 entry = huge_ptep_get(ptep);
2648                 if (unlikely(is_hugetlb_entry_migration(entry))) {
2649                         migration_entry_wait(mm, (pmd_t *)ptep, address);
2650                         return 0;
2651                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2652                         return VM_FAULT_HWPOISON_LARGE |
2653                                VM_FAULT_SET_HINDEX(h - hstates);
2654         }
2655
2656         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2657         if (!ptep)
2658                 return VM_FAULT_OOM;
2659
2660         /*
2661          * Serialize hugepage allocation and instantiation, so that we don't
2662          * get spurious allocation failures if two CPUs race to instantiate
2663          * the same page in the page cache.
2664          */
2665         mutex_lock(&hugetlb_instantiation_mutex);
2666         entry = huge_ptep_get(ptep);
2667         if (huge_pte_none(entry)) {
2668                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2669                 goto out_mutex;
2670         }
2671
2672         ret = 0;
2673
2674         /*
2675          * If we are going to COW the mapping later, we examine the pending
2676          * reservations for this page now. This will ensure that any
2677          * allocations necessary to record that reservation occur outside the
2678          * spinlock. For private mappings, we also lookup the pagecache
2679          * page now as it is used to determine if a reservation has been
2680          * consumed.
2681          */
2682         if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2683                 if (vma_needs_reservation(h, vma, address) < 0) {
2684                         ret = VM_FAULT_OOM;
2685                         goto out_mutex;
2686                 }
2687
2688                 if (!(vma->vm_flags & VM_MAYSHARE))
2689                         pagecache_page = hugetlbfs_pagecache_page(h,
2690                                                                 vma, address);
2691         }
2692
2693         /*
2694          * hugetlb_cow() requires page locks of pte_page(entry) and
2695          * pagecache_page, so here we need take the former one
2696          * when page != pagecache_page or !pagecache_page.
2697          * Note that locking order is always pagecache_page -> page,
2698          * so no worry about deadlock.
2699          */
2700         page = pte_page(entry);
2701         if (page != pagecache_page)
2702                 lock_page(page);
2703
2704         spin_lock(&mm->page_table_lock);
2705         /* Check for a racing update before calling hugetlb_cow */
2706         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2707                 goto out_page_table_lock;
2708
2709
2710         if (flags & FAULT_FLAG_WRITE) {
2711                 if (!pte_write(entry)) {
2712                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2713                                                         pagecache_page);
2714                         goto out_page_table_lock;
2715                 }
2716                 entry = pte_mkdirty(entry);
2717         }
2718         entry = pte_mkyoung(entry);
2719         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2720                                                 flags & FAULT_FLAG_WRITE))
2721                 update_mmu_cache(vma, address, ptep);
2722
2723 out_page_table_lock:
2724         spin_unlock(&mm->page_table_lock);
2725
2726         if (pagecache_page) {
2727                 unlock_page(pagecache_page);
2728                 put_page(pagecache_page);
2729         }
2730         if (page != pagecache_page)
2731                 unlock_page(page);
2732
2733 out_mutex:
2734         mutex_unlock(&hugetlb_instantiation_mutex);
2735
2736         return ret;
2737 }
2738
2739 /* Can be overriden by architectures */
2740 __attribute__((weak)) struct page *
2741 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2742                pud_t *pud, int write)
2743 {
2744         BUG();
2745         return NULL;
2746 }
2747
2748 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2749                         struct page **pages, struct vm_area_struct **vmas,
2750                         unsigned long *position, int *length, int i,
2751                         unsigned int flags)
2752 {
2753         unsigned long pfn_offset;
2754         unsigned long vaddr = *position;
2755         int remainder = *length;
2756         struct hstate *h = hstate_vma(vma);
2757
2758         spin_lock(&mm->page_table_lock);
2759         while (vaddr < vma->vm_end && remainder) {
2760                 pte_t *pte;
2761                 int absent;
2762                 struct page *page;
2763
2764                 /*
2765                  * Some archs (sparc64, sh*) have multiple pte_ts to
2766                  * each hugepage.  We have to make sure we get the
2767                  * first, for the page indexing below to work.
2768                  */
2769                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2770                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2771
2772                 /*
2773                  * When coredumping, it suits get_dump_page if we just return
2774                  * an error where there's an empty slot with no huge pagecache
2775                  * to back it.  This way, we avoid allocating a hugepage, and
2776                  * the sparse dumpfile avoids allocating disk blocks, but its
2777                  * huge holes still show up with zeroes where they need to be.
2778                  */
2779                 if (absent && (flags & FOLL_DUMP) &&
2780                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2781                         remainder = 0;
2782                         break;
2783                 }
2784
2785                 if (absent ||
2786                     ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2787                         int ret;
2788
2789                         spin_unlock(&mm->page_table_lock);
2790                         ret = hugetlb_fault(mm, vma, vaddr,
2791                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2792                         spin_lock(&mm->page_table_lock);
2793                         if (!(ret & VM_FAULT_ERROR))
2794                                 continue;
2795
2796                         remainder = 0;
2797                         break;
2798                 }
2799
2800                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2801                 page = pte_page(huge_ptep_get(pte));
2802 same_page:
2803                 if (pages) {
2804                         pages[i] = mem_map_offset(page, pfn_offset);
2805                         get_page(pages[i]);
2806                 }
2807
2808                 if (vmas)
2809                         vmas[i] = vma;
2810
2811                 vaddr += PAGE_SIZE;
2812                 ++pfn_offset;
2813                 --remainder;
2814                 ++i;
2815                 if (vaddr < vma->vm_end && remainder &&
2816                                 pfn_offset < pages_per_huge_page(h)) {
2817                         /*
2818                          * We use pfn_offset to avoid touching the pageframes
2819                          * of this compound page.
2820                          */
2821                         goto same_page;
2822                 }
2823         }
2824         spin_unlock(&mm->page_table_lock);
2825         *length = remainder;
2826         *position = vaddr;
2827
2828         return i ? i : -EFAULT;
2829 }
2830
2831 void hugetlb_change_protection(struct vm_area_struct *vma,
2832                 unsigned long address, unsigned long end, pgprot_t newprot)
2833 {
2834         struct mm_struct *mm = vma->vm_mm;
2835         unsigned long start = address;
2836         pte_t *ptep;
2837         pte_t pte;
2838         struct hstate *h = hstate_vma(vma);
2839
2840         BUG_ON(address >= end);
2841         flush_cache_range(vma, address, end);
2842
2843         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2844         spin_lock(&mm->page_table_lock);
2845         for (; address < end; address += huge_page_size(h)) {
2846                 ptep = huge_pte_offset(mm, address);
2847                 if (!ptep)
2848                         continue;
2849                 if (huge_pmd_unshare(mm, &address, ptep))
2850                         continue;
2851                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2852                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2853                         pte = pte_mkhuge(pte_modify(pte, newprot));
2854                         set_huge_pte_at(mm, address, ptep, pte);
2855                 }
2856         }
2857         spin_unlock(&mm->page_table_lock);
2858         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2859
2860         flush_tlb_range(vma, start, end);
2861 }
2862
2863 int hugetlb_reserve_pages(struct inode *inode,
2864                                         long from, long to,
2865                                         struct vm_area_struct *vma,
2866                                         vm_flags_t vm_flags)
2867 {
2868         long ret, chg;
2869         struct hstate *h = hstate_inode(inode);
2870
2871         /*
2872          * Only apply hugepage reservation if asked. At fault time, an
2873          * attempt will be made for VM_NORESERVE to allocate a page
2874          * and filesystem quota without using reserves
2875          */
2876         if (vm_flags & VM_NORESERVE)
2877                 return 0;
2878
2879         /*
2880          * Shared mappings base their reservation on the number of pages that
2881          * are already allocated on behalf of the file. Private mappings need
2882          * to reserve the full area even if read-only as mprotect() may be
2883          * called to make the mapping read-write. Assume !vma is a shm mapping
2884          */
2885         if (!vma || vma->vm_flags & VM_MAYSHARE)
2886                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2887         else {
2888                 struct resv_map *resv_map = resv_map_alloc();
2889                 if (!resv_map)
2890                         return -ENOMEM;
2891
2892                 chg = to - from;
2893
2894                 set_vma_resv_map(vma, resv_map);
2895                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2896         }
2897
2898         if (chg < 0)
2899                 return chg;
2900
2901         /* There must be enough filesystem quota for the mapping */
2902         if (hugetlb_get_quota(inode->i_mapping, chg))
2903                 return -ENOSPC;
2904
2905         /*
2906          * Check enough hugepages are available for the reservation.
2907          * Hand back the quota if there are not
2908          */
2909         ret = hugetlb_acct_memory(h, chg);
2910         if (ret < 0) {
2911                 hugetlb_put_quota(inode->i_mapping, chg);
2912                 return ret;
2913         }
2914
2915         /*
2916          * Account for the reservations made. Shared mappings record regions
2917          * that have reservations as they are shared by multiple VMAs.
2918          * When the last VMA disappears, the region map says how much
2919          * the reservation was and the page cache tells how much of
2920          * the reservation was consumed. Private mappings are per-VMA and
2921          * only the consumed reservations are tracked. When the VMA
2922          * disappears, the original reservation is the VMA size and the
2923          * consumed reservations are stored in the map. Hence, nothing
2924          * else has to be done for private mappings here
2925          */
2926         if (!vma || vma->vm_flags & VM_MAYSHARE)
2927                 region_add(&inode->i_mapping->private_list, from, to);
2928         return 0;
2929 }
2930
2931 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2932 {
2933         struct hstate *h = hstate_inode(inode);
2934         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2935
2936         spin_lock(&inode->i_lock);
2937         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2938         spin_unlock(&inode->i_lock);
2939
2940         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2941         hugetlb_acct_memory(h, -(chg - freed));
2942 }
2943
2944 #ifdef CONFIG_MEMORY_FAILURE
2945
2946 /* Should be called in hugetlb_lock */
2947 static int is_hugepage_on_freelist(struct page *hpage)
2948 {
2949         struct page *page;
2950         struct page *tmp;
2951         struct hstate *h = page_hstate(hpage);
2952         int nid = page_to_nid(hpage);
2953
2954         list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
2955                 if (page == hpage)
2956                         return 1;
2957         return 0;
2958 }
2959
2960 /*
2961  * This function is called from memory failure code.
2962  * Assume the caller holds page lock of the head page.
2963  */
2964 int dequeue_hwpoisoned_huge_page(struct page *hpage)
2965 {
2966         struct hstate *h = page_hstate(hpage);
2967         int nid = page_to_nid(hpage);
2968         int ret = -EBUSY;
2969
2970         spin_lock(&hugetlb_lock);
2971         if (is_hugepage_on_freelist(hpage)) {
2972                 list_del(&hpage->lru);
2973                 set_page_refcounted(hpage);
2974                 h->free_huge_pages--;
2975                 h->free_huge_pages_node[nid]--;
2976                 ret = 0;
2977         }
2978         spin_unlock(&hugetlb_lock);
2979         return ret;
2980 }
2981 #endif