Merge branch 'akpm-incoming-1'
[pandora-kernel.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 <asm/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 futher 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 clear_gigantic_page(struct page *page,
398                         unsigned long addr, unsigned long sz)
399 {
400         int i;
401         struct page *p = page;
402
403         might_sleep();
404         for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
405                 cond_resched();
406                 clear_user_highpage(p, addr + i * PAGE_SIZE);
407         }
408 }
409 static void clear_huge_page(struct page *page,
410                         unsigned long addr, unsigned long sz)
411 {
412         int i;
413
414         if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) {
415                 clear_gigantic_page(page, addr, sz);
416                 return;
417         }
418
419         might_sleep();
420         for (i = 0; i < sz/PAGE_SIZE; i++) {
421                 cond_resched();
422                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
423         }
424 }
425
426 static void copy_user_gigantic_page(struct page *dst, struct page *src,
427                            unsigned long addr, struct vm_area_struct *vma)
428 {
429         int i;
430         struct hstate *h = hstate_vma(vma);
431         struct page *dst_base = dst;
432         struct page *src_base = src;
433
434         for (i = 0; i < pages_per_huge_page(h); ) {
435                 cond_resched();
436                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
437
438                 i++;
439                 dst = mem_map_next(dst, dst_base, i);
440                 src = mem_map_next(src, src_base, i);
441         }
442 }
443
444 static void copy_user_huge_page(struct page *dst, struct page *src,
445                            unsigned long addr, struct vm_area_struct *vma)
446 {
447         int i;
448         struct hstate *h = hstate_vma(vma);
449
450         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
451                 copy_user_gigantic_page(dst, src, addr, vma);
452                 return;
453         }
454
455         might_sleep();
456         for (i = 0; i < pages_per_huge_page(h); i++) {
457                 cond_resched();
458                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
459         }
460 }
461
462 static void copy_gigantic_page(struct page *dst, struct page *src)
463 {
464         int i;
465         struct hstate *h = page_hstate(src);
466         struct page *dst_base = dst;
467         struct page *src_base = src;
468
469         for (i = 0; i < pages_per_huge_page(h); ) {
470                 cond_resched();
471                 copy_highpage(dst, src);
472
473                 i++;
474                 dst = mem_map_next(dst, dst_base, i);
475                 src = mem_map_next(src, src_base, i);
476         }
477 }
478
479 void copy_huge_page(struct page *dst, struct page *src)
480 {
481         int i;
482         struct hstate *h = page_hstate(src);
483
484         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
485                 copy_gigantic_page(dst, src);
486                 return;
487         }
488
489         might_sleep();
490         for (i = 0; i < pages_per_huge_page(h); i++) {
491                 cond_resched();
492                 copy_highpage(dst + i, src + i);
493         }
494 }
495
496 static void enqueue_huge_page(struct hstate *h, struct page *page)
497 {
498         int nid = page_to_nid(page);
499         list_add(&page->lru, &h->hugepage_freelists[nid]);
500         h->free_huge_pages++;
501         h->free_huge_pages_node[nid]++;
502 }
503
504 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
505 {
506         struct page *page;
507
508         if (list_empty(&h->hugepage_freelists[nid]))
509                 return NULL;
510         page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
511         list_del(&page->lru);
512         set_page_refcounted(page);
513         h->free_huge_pages--;
514         h->free_huge_pages_node[nid]--;
515         return page;
516 }
517
518 static struct page *dequeue_huge_page_vma(struct hstate *h,
519                                 struct vm_area_struct *vma,
520                                 unsigned long address, int avoid_reserve)
521 {
522         struct page *page = NULL;
523         struct mempolicy *mpol;
524         nodemask_t *nodemask;
525         struct zonelist *zonelist;
526         struct zone *zone;
527         struct zoneref *z;
528
529         get_mems_allowed();
530         zonelist = huge_zonelist(vma, address,
531                                         htlb_alloc_mask, &mpol, &nodemask);
532         /*
533          * A child process with MAP_PRIVATE mappings created by their parent
534          * have no page reserves. This check ensures that reservations are
535          * not "stolen". The child may still get SIGKILLed
536          */
537         if (!vma_has_reserves(vma) &&
538                         h->free_huge_pages - h->resv_huge_pages == 0)
539                 goto err;
540
541         /* If reserves cannot be used, ensure enough pages are in the pool */
542         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
543                 goto err;;
544
545         for_each_zone_zonelist_nodemask(zone, z, zonelist,
546                                                 MAX_NR_ZONES - 1, nodemask) {
547                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
548                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
549                         if (page) {
550                                 if (!avoid_reserve)
551                                         decrement_hugepage_resv_vma(h, vma);
552                                 break;
553                         }
554                 }
555         }
556 err:
557         mpol_cond_put(mpol);
558         put_mems_allowed();
559         return page;
560 }
561
562 static void update_and_free_page(struct hstate *h, struct page *page)
563 {
564         int i;
565
566         VM_BUG_ON(h->order >= MAX_ORDER);
567
568         h->nr_huge_pages--;
569         h->nr_huge_pages_node[page_to_nid(page)]--;
570         for (i = 0; i < pages_per_huge_page(h); i++) {
571                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
572                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
573                                 1 << PG_private | 1<< PG_writeback);
574         }
575         set_compound_page_dtor(page, NULL);
576         set_page_refcounted(page);
577         arch_release_hugepage(page);
578         __free_pages(page, huge_page_order(h));
579 }
580
581 struct hstate *size_to_hstate(unsigned long size)
582 {
583         struct hstate *h;
584
585         for_each_hstate(h) {
586                 if (huge_page_size(h) == size)
587                         return h;
588         }
589         return NULL;
590 }
591
592 static void free_huge_page(struct page *page)
593 {
594         /*
595          * Can't pass hstate in here because it is called from the
596          * compound page destructor.
597          */
598         struct hstate *h = page_hstate(page);
599         int nid = page_to_nid(page);
600         struct address_space *mapping;
601
602         mapping = (struct address_space *) page_private(page);
603         set_page_private(page, 0);
604         page->mapping = NULL;
605         BUG_ON(page_count(page));
606         BUG_ON(page_mapcount(page));
607         INIT_LIST_HEAD(&page->lru);
608
609         spin_lock(&hugetlb_lock);
610         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
611                 update_and_free_page(h, page);
612                 h->surplus_huge_pages--;
613                 h->surplus_huge_pages_node[nid]--;
614         } else {
615                 enqueue_huge_page(h, page);
616         }
617         spin_unlock(&hugetlb_lock);
618         if (mapping)
619                 hugetlb_put_quota(mapping, 1);
620 }
621
622 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
623 {
624         set_compound_page_dtor(page, free_huge_page);
625         spin_lock(&hugetlb_lock);
626         h->nr_huge_pages++;
627         h->nr_huge_pages_node[nid]++;
628         spin_unlock(&hugetlb_lock);
629         put_page(page); /* free it into the hugepage allocator */
630 }
631
632 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
633 {
634         int i;
635         int nr_pages = 1 << order;
636         struct page *p = page + 1;
637
638         /* we rely on prep_new_huge_page to set the destructor */
639         set_compound_order(page, order);
640         __SetPageHead(page);
641         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
642                 __SetPageTail(p);
643                 p->first_page = page;
644         }
645 }
646
647 int PageHuge(struct page *page)
648 {
649         compound_page_dtor *dtor;
650
651         if (!PageCompound(page))
652                 return 0;
653
654         page = compound_head(page);
655         dtor = get_compound_page_dtor(page);
656
657         return dtor == free_huge_page;
658 }
659
660 EXPORT_SYMBOL_GPL(PageHuge);
661
662 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
663 {
664         struct page *page;
665
666         if (h->order >= MAX_ORDER)
667                 return NULL;
668
669         page = alloc_pages_exact_node(nid,
670                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
671                                                 __GFP_REPEAT|__GFP_NOWARN,
672                 huge_page_order(h));
673         if (page) {
674                 if (arch_prepare_hugepage(page)) {
675                         __free_pages(page, huge_page_order(h));
676                         return NULL;
677                 }
678                 prep_new_huge_page(h, page, nid);
679         }
680
681         return page;
682 }
683
684 /*
685  * common helper functions for hstate_next_node_to_{alloc|free}.
686  * We may have allocated or freed a huge page based on a different
687  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
688  * be outside of *nodes_allowed.  Ensure that we use an allowed
689  * node for alloc or free.
690  */
691 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
692 {
693         nid = next_node(nid, *nodes_allowed);
694         if (nid == MAX_NUMNODES)
695                 nid = first_node(*nodes_allowed);
696         VM_BUG_ON(nid >= MAX_NUMNODES);
697
698         return nid;
699 }
700
701 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
702 {
703         if (!node_isset(nid, *nodes_allowed))
704                 nid = next_node_allowed(nid, nodes_allowed);
705         return nid;
706 }
707
708 /*
709  * returns the previously saved node ["this node"] from which to
710  * allocate a persistent huge page for the pool and advance the
711  * next node from which to allocate, handling wrap at end of node
712  * mask.
713  */
714 static int hstate_next_node_to_alloc(struct hstate *h,
715                                         nodemask_t *nodes_allowed)
716 {
717         int nid;
718
719         VM_BUG_ON(!nodes_allowed);
720
721         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
722         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
723
724         return nid;
725 }
726
727 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
728 {
729         struct page *page;
730         int start_nid;
731         int next_nid;
732         int ret = 0;
733
734         start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
735         next_nid = start_nid;
736
737         do {
738                 page = alloc_fresh_huge_page_node(h, next_nid);
739                 if (page) {
740                         ret = 1;
741                         break;
742                 }
743                 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
744         } while (next_nid != start_nid);
745
746         if (ret)
747                 count_vm_event(HTLB_BUDDY_PGALLOC);
748         else
749                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
750
751         return ret;
752 }
753
754 /*
755  * helper for free_pool_huge_page() - return the previously saved
756  * node ["this node"] from which to free a huge page.  Advance the
757  * next node id whether or not we find a free huge page to free so
758  * that the next attempt to free addresses the next node.
759  */
760 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
761 {
762         int nid;
763
764         VM_BUG_ON(!nodes_allowed);
765
766         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
767         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
768
769         return nid;
770 }
771
772 /*
773  * Free huge page from pool from next node to free.
774  * Attempt to keep persistent huge pages more or less
775  * balanced over allowed nodes.
776  * Called with hugetlb_lock locked.
777  */
778 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
779                                                          bool acct_surplus)
780 {
781         int start_nid;
782         int next_nid;
783         int ret = 0;
784
785         start_nid = hstate_next_node_to_free(h, nodes_allowed);
786         next_nid = start_nid;
787
788         do {
789                 /*
790                  * If we're returning unused surplus pages, only examine
791                  * nodes with surplus pages.
792                  */
793                 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
794                     !list_empty(&h->hugepage_freelists[next_nid])) {
795                         struct page *page =
796                                 list_entry(h->hugepage_freelists[next_nid].next,
797                                           struct page, lru);
798                         list_del(&page->lru);
799                         h->free_huge_pages--;
800                         h->free_huge_pages_node[next_nid]--;
801                         if (acct_surplus) {
802                                 h->surplus_huge_pages--;
803                                 h->surplus_huge_pages_node[next_nid]--;
804                         }
805                         update_and_free_page(h, page);
806                         ret = 1;
807                         break;
808                 }
809                 next_nid = hstate_next_node_to_free(h, nodes_allowed);
810         } while (next_nid != start_nid);
811
812         return ret;
813 }
814
815 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
816 {
817         struct page *page;
818         unsigned int r_nid;
819
820         if (h->order >= MAX_ORDER)
821                 return NULL;
822
823         /*
824          * Assume we will successfully allocate the surplus page to
825          * prevent racing processes from causing the surplus to exceed
826          * overcommit
827          *
828          * This however introduces a different race, where a process B
829          * tries to grow the static hugepage pool while alloc_pages() is
830          * called by process A. B will only examine the per-node
831          * counters in determining if surplus huge pages can be
832          * converted to normal huge pages in adjust_pool_surplus(). A
833          * won't be able to increment the per-node counter, until the
834          * lock is dropped by B, but B doesn't drop hugetlb_lock until
835          * no more huge pages can be converted from surplus to normal
836          * state (and doesn't try to convert again). Thus, we have a
837          * case where a surplus huge page exists, the pool is grown, and
838          * the surplus huge page still exists after, even though it
839          * should just have been converted to a normal huge page. This
840          * does not leak memory, though, as the hugepage will be freed
841          * once it is out of use. It also does not allow the counters to
842          * go out of whack in adjust_pool_surplus() as we don't modify
843          * the node values until we've gotten the hugepage and only the
844          * per-node value is checked there.
845          */
846         spin_lock(&hugetlb_lock);
847         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
848                 spin_unlock(&hugetlb_lock);
849                 return NULL;
850         } else {
851                 h->nr_huge_pages++;
852                 h->surplus_huge_pages++;
853         }
854         spin_unlock(&hugetlb_lock);
855
856         if (nid == NUMA_NO_NODE)
857                 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
858                                    __GFP_REPEAT|__GFP_NOWARN,
859                                    huge_page_order(h));
860         else
861                 page = alloc_pages_exact_node(nid,
862                         htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
863                         __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
864
865         if (page && arch_prepare_hugepage(page)) {
866                 __free_pages(page, huge_page_order(h));
867                 return NULL;
868         }
869
870         spin_lock(&hugetlb_lock);
871         if (page) {
872                 r_nid = page_to_nid(page);
873                 set_compound_page_dtor(page, free_huge_page);
874                 /*
875                  * We incremented the global counters already
876                  */
877                 h->nr_huge_pages_node[r_nid]++;
878                 h->surplus_huge_pages_node[r_nid]++;
879                 __count_vm_event(HTLB_BUDDY_PGALLOC);
880         } else {
881                 h->nr_huge_pages--;
882                 h->surplus_huge_pages--;
883                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
884         }
885         spin_unlock(&hugetlb_lock);
886
887         return page;
888 }
889
890 /*
891  * This allocation function is useful in the context where vma is irrelevant.
892  * E.g. soft-offlining uses this function because it only cares physical
893  * address of error page.
894  */
895 struct page *alloc_huge_page_node(struct hstate *h, int nid)
896 {
897         struct page *page;
898
899         spin_lock(&hugetlb_lock);
900         page = dequeue_huge_page_node(h, nid);
901         spin_unlock(&hugetlb_lock);
902
903         if (!page)
904                 page = alloc_buddy_huge_page(h, nid);
905
906         return page;
907 }
908
909 /*
910  * Increase the hugetlb pool such that it can accomodate a reservation
911  * of size 'delta'.
912  */
913 static int gather_surplus_pages(struct hstate *h, int delta)
914 {
915         struct list_head surplus_list;
916         struct page *page, *tmp;
917         int ret, i;
918         int needed, allocated;
919
920         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
921         if (needed <= 0) {
922                 h->resv_huge_pages += delta;
923                 return 0;
924         }
925
926         allocated = 0;
927         INIT_LIST_HEAD(&surplus_list);
928
929         ret = -ENOMEM;
930 retry:
931         spin_unlock(&hugetlb_lock);
932         for (i = 0; i < needed; i++) {
933                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
934                 if (!page)
935                         /*
936                          * We were not able to allocate enough pages to
937                          * satisfy the entire reservation so we free what
938                          * we've allocated so far.
939                          */
940                         goto free;
941
942                 list_add(&page->lru, &surplus_list);
943         }
944         allocated += needed;
945
946         /*
947          * After retaking hugetlb_lock, we need to recalculate 'needed'
948          * because either resv_huge_pages or free_huge_pages may have changed.
949          */
950         spin_lock(&hugetlb_lock);
951         needed = (h->resv_huge_pages + delta) -
952                         (h->free_huge_pages + allocated);
953         if (needed > 0)
954                 goto retry;
955
956         /*
957          * The surplus_list now contains _at_least_ the number of extra pages
958          * needed to accomodate the reservation.  Add the appropriate number
959          * of pages to the hugetlb pool and free the extras back to the buddy
960          * allocator.  Commit the entire reservation here to prevent another
961          * process from stealing the pages as they are added to the pool but
962          * before they are reserved.
963          */
964         needed += allocated;
965         h->resv_huge_pages += delta;
966         ret = 0;
967
968         spin_unlock(&hugetlb_lock);
969         /* Free the needed pages to the hugetlb pool */
970         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
971                 if ((--needed) < 0)
972                         break;
973                 list_del(&page->lru);
974                 /*
975                  * This page is now managed by the hugetlb allocator and has
976                  * no users -- drop the buddy allocator's reference.
977                  */
978                 put_page_testzero(page);
979                 VM_BUG_ON(page_count(page));
980                 enqueue_huge_page(h, page);
981         }
982
983         /* Free unnecessary surplus pages to the buddy allocator */
984 free:
985         if (!list_empty(&surplus_list)) {
986                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
987                         list_del(&page->lru);
988                         put_page(page);
989                 }
990         }
991         spin_lock(&hugetlb_lock);
992
993         return ret;
994 }
995
996 /*
997  * When releasing a hugetlb pool reservation, any surplus pages that were
998  * allocated to satisfy the reservation must be explicitly freed if they were
999  * never used.
1000  * Called with hugetlb_lock held.
1001  */
1002 static void return_unused_surplus_pages(struct hstate *h,
1003                                         unsigned long unused_resv_pages)
1004 {
1005         unsigned long nr_pages;
1006
1007         /* Uncommit the reservation */
1008         h->resv_huge_pages -= unused_resv_pages;
1009
1010         /* Cannot return gigantic pages currently */
1011         if (h->order >= MAX_ORDER)
1012                 return;
1013
1014         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1015
1016         /*
1017          * We want to release as many surplus pages as possible, spread
1018          * evenly across all nodes with memory. Iterate across these nodes
1019          * until we can no longer free unreserved surplus pages. This occurs
1020          * when the nodes with surplus pages have no free pages.
1021          * free_pool_huge_page() will balance the the freed pages across the
1022          * on-line nodes with memory and will handle the hstate accounting.
1023          */
1024         while (nr_pages--) {
1025                 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1026                         break;
1027         }
1028 }
1029
1030 /*
1031  * Determine if the huge page at addr within the vma has an associated
1032  * reservation.  Where it does not we will need to logically increase
1033  * reservation and actually increase quota before an allocation can occur.
1034  * Where any new reservation would be required the reservation change is
1035  * prepared, but not committed.  Once the page has been quota'd allocated
1036  * an instantiated the change should be committed via vma_commit_reservation.
1037  * No action is required on failure.
1038  */
1039 static long vma_needs_reservation(struct hstate *h,
1040                         struct vm_area_struct *vma, unsigned long addr)
1041 {
1042         struct address_space *mapping = vma->vm_file->f_mapping;
1043         struct inode *inode = mapping->host;
1044
1045         if (vma->vm_flags & VM_MAYSHARE) {
1046                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1047                 return region_chg(&inode->i_mapping->private_list,
1048                                                         idx, idx + 1);
1049
1050         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1051                 return 1;
1052
1053         } else  {
1054                 long err;
1055                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1056                 struct resv_map *reservations = vma_resv_map(vma);
1057
1058                 err = region_chg(&reservations->regions, idx, idx + 1);
1059                 if (err < 0)
1060                         return err;
1061                 return 0;
1062         }
1063 }
1064 static void vma_commit_reservation(struct hstate *h,
1065                         struct vm_area_struct *vma, unsigned long addr)
1066 {
1067         struct address_space *mapping = vma->vm_file->f_mapping;
1068         struct inode *inode = mapping->host;
1069
1070         if (vma->vm_flags & VM_MAYSHARE) {
1071                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1072                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1073
1074         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1075                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1076                 struct resv_map *reservations = vma_resv_map(vma);
1077
1078                 /* Mark this page used in the map. */
1079                 region_add(&reservations->regions, idx, idx + 1);
1080         }
1081 }
1082
1083 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1084                                     unsigned long addr, int avoid_reserve)
1085 {
1086         struct hstate *h = hstate_vma(vma);
1087         struct page *page;
1088         struct address_space *mapping = vma->vm_file->f_mapping;
1089         struct inode *inode = mapping->host;
1090         long chg;
1091
1092         /*
1093          * Processes that did not create the mapping will have no reserves and
1094          * will not have accounted against quota. Check that the quota can be
1095          * made before satisfying the allocation
1096          * MAP_NORESERVE mappings may also need pages and quota allocated
1097          * if no reserve mapping overlaps.
1098          */
1099         chg = vma_needs_reservation(h, vma, addr);
1100         if (chg < 0)
1101                 return ERR_PTR(chg);
1102         if (chg)
1103                 if (hugetlb_get_quota(inode->i_mapping, chg))
1104                         return ERR_PTR(-ENOSPC);
1105
1106         spin_lock(&hugetlb_lock);
1107         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1108         spin_unlock(&hugetlb_lock);
1109
1110         if (!page) {
1111                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1112                 if (!page) {
1113                         hugetlb_put_quota(inode->i_mapping, chg);
1114                         return ERR_PTR(-VM_FAULT_SIGBUS);
1115                 }
1116         }
1117
1118         set_page_private(page, (unsigned long) mapping);
1119
1120         vma_commit_reservation(h, vma, addr);
1121
1122         return page;
1123 }
1124
1125 int __weak alloc_bootmem_huge_page(struct hstate *h)
1126 {
1127         struct huge_bootmem_page *m;
1128         int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1129
1130         while (nr_nodes) {
1131                 void *addr;
1132
1133                 addr = __alloc_bootmem_node_nopanic(
1134                                 NODE_DATA(hstate_next_node_to_alloc(h,
1135                                                 &node_states[N_HIGH_MEMORY])),
1136                                 huge_page_size(h), huge_page_size(h), 0);
1137
1138                 if (addr) {
1139                         /*
1140                          * Use the beginning of the huge page to store the
1141                          * huge_bootmem_page struct (until gather_bootmem
1142                          * puts them into the mem_map).
1143                          */
1144                         m = addr;
1145                         goto found;
1146                 }
1147                 nr_nodes--;
1148         }
1149         return 0;
1150
1151 found:
1152         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1153         /* Put them into a private list first because mem_map is not up yet */
1154         list_add(&m->list, &huge_boot_pages);
1155         m->hstate = h;
1156         return 1;
1157 }
1158
1159 static void prep_compound_huge_page(struct page *page, int order)
1160 {
1161         if (unlikely(order > (MAX_ORDER - 1)))
1162                 prep_compound_gigantic_page(page, order);
1163         else
1164                 prep_compound_page(page, order);
1165 }
1166
1167 /* Put bootmem huge pages into the standard lists after mem_map is up */
1168 static void __init gather_bootmem_prealloc(void)
1169 {
1170         struct huge_bootmem_page *m;
1171
1172         list_for_each_entry(m, &huge_boot_pages, list) {
1173                 struct page *page = virt_to_page(m);
1174                 struct hstate *h = m->hstate;
1175                 __ClearPageReserved(page);
1176                 WARN_ON(page_count(page) != 1);
1177                 prep_compound_huge_page(page, h->order);
1178                 prep_new_huge_page(h, page, page_to_nid(page));
1179         }
1180 }
1181
1182 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1183 {
1184         unsigned long i;
1185
1186         for (i = 0; i < h->max_huge_pages; ++i) {
1187                 if (h->order >= MAX_ORDER) {
1188                         if (!alloc_bootmem_huge_page(h))
1189                                 break;
1190                 } else if (!alloc_fresh_huge_page(h,
1191                                          &node_states[N_HIGH_MEMORY]))
1192                         break;
1193         }
1194         h->max_huge_pages = i;
1195 }
1196
1197 static void __init hugetlb_init_hstates(void)
1198 {
1199         struct hstate *h;
1200
1201         for_each_hstate(h) {
1202                 /* oversize hugepages were init'ed in early boot */
1203                 if (h->order < MAX_ORDER)
1204                         hugetlb_hstate_alloc_pages(h);
1205         }
1206 }
1207
1208 static char * __init memfmt(char *buf, unsigned long n)
1209 {
1210         if (n >= (1UL << 30))
1211                 sprintf(buf, "%lu GB", n >> 30);
1212         else if (n >= (1UL << 20))
1213                 sprintf(buf, "%lu MB", n >> 20);
1214         else
1215                 sprintf(buf, "%lu KB", n >> 10);
1216         return buf;
1217 }
1218
1219 static void __init report_hugepages(void)
1220 {
1221         struct hstate *h;
1222
1223         for_each_hstate(h) {
1224                 char buf[32];
1225                 printk(KERN_INFO "HugeTLB registered %s page size, "
1226                                  "pre-allocated %ld pages\n",
1227                         memfmt(buf, huge_page_size(h)),
1228                         h->free_huge_pages);
1229         }
1230 }
1231
1232 #ifdef CONFIG_HIGHMEM
1233 static void try_to_free_low(struct hstate *h, unsigned long count,
1234                                                 nodemask_t *nodes_allowed)
1235 {
1236         int i;
1237
1238         if (h->order >= MAX_ORDER)
1239                 return;
1240
1241         for_each_node_mask(i, *nodes_allowed) {
1242                 struct page *page, *next;
1243                 struct list_head *freel = &h->hugepage_freelists[i];
1244                 list_for_each_entry_safe(page, next, freel, lru) {
1245                         if (count >= h->nr_huge_pages)
1246                                 return;
1247                         if (PageHighMem(page))
1248                                 continue;
1249                         list_del(&page->lru);
1250                         update_and_free_page(h, page);
1251                         h->free_huge_pages--;
1252                         h->free_huge_pages_node[page_to_nid(page)]--;
1253                 }
1254         }
1255 }
1256 #else
1257 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1258                                                 nodemask_t *nodes_allowed)
1259 {
1260 }
1261 #endif
1262
1263 /*
1264  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1265  * balanced by operating on them in a round-robin fashion.
1266  * Returns 1 if an adjustment was made.
1267  */
1268 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1269                                 int delta)
1270 {
1271         int start_nid, next_nid;
1272         int ret = 0;
1273
1274         VM_BUG_ON(delta != -1 && delta != 1);
1275
1276         if (delta < 0)
1277                 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1278         else
1279                 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1280         next_nid = start_nid;
1281
1282         do {
1283                 int nid = next_nid;
1284                 if (delta < 0)  {
1285                         /*
1286                          * To shrink on this node, there must be a surplus page
1287                          */
1288                         if (!h->surplus_huge_pages_node[nid]) {
1289                                 next_nid = hstate_next_node_to_alloc(h,
1290                                                                 nodes_allowed);
1291                                 continue;
1292                         }
1293                 }
1294                 if (delta > 0) {
1295                         /*
1296                          * Surplus cannot exceed the total number of pages
1297                          */
1298                         if (h->surplus_huge_pages_node[nid] >=
1299                                                 h->nr_huge_pages_node[nid]) {
1300                                 next_nid = hstate_next_node_to_free(h,
1301                                                                 nodes_allowed);
1302                                 continue;
1303                         }
1304                 }
1305
1306                 h->surplus_huge_pages += delta;
1307                 h->surplus_huge_pages_node[nid] += delta;
1308                 ret = 1;
1309                 break;
1310         } while (next_nid != start_nid);
1311
1312         return ret;
1313 }
1314
1315 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1316 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1317                                                 nodemask_t *nodes_allowed)
1318 {
1319         unsigned long min_count, ret;
1320
1321         if (h->order >= MAX_ORDER)
1322                 return h->max_huge_pages;
1323
1324         /*
1325          * Increase the pool size
1326          * First take pages out of surplus state.  Then make up the
1327          * remaining difference by allocating fresh huge pages.
1328          *
1329          * We might race with alloc_buddy_huge_page() here and be unable
1330          * to convert a surplus huge page to a normal huge page. That is
1331          * not critical, though, it just means the overall size of the
1332          * pool might be one hugepage larger than it needs to be, but
1333          * within all the constraints specified by the sysctls.
1334          */
1335         spin_lock(&hugetlb_lock);
1336         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1337                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1338                         break;
1339         }
1340
1341         while (count > persistent_huge_pages(h)) {
1342                 /*
1343                  * If this allocation races such that we no longer need the
1344                  * page, free_huge_page will handle it by freeing the page
1345                  * and reducing the surplus.
1346                  */
1347                 spin_unlock(&hugetlb_lock);
1348                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1349                 spin_lock(&hugetlb_lock);
1350                 if (!ret)
1351                         goto out;
1352
1353                 /* Bail for signals. Probably ctrl-c from user */
1354                 if (signal_pending(current))
1355                         goto out;
1356         }
1357
1358         /*
1359          * Decrease the pool size
1360          * First return free pages to the buddy allocator (being careful
1361          * to keep enough around to satisfy reservations).  Then place
1362          * pages into surplus state as needed so the pool will shrink
1363          * to the desired size as pages become free.
1364          *
1365          * By placing pages into the surplus state independent of the
1366          * overcommit value, we are allowing the surplus pool size to
1367          * exceed overcommit. There are few sane options here. Since
1368          * alloc_buddy_huge_page() is checking the global counter,
1369          * though, we'll note that we're not allowed to exceed surplus
1370          * and won't grow the pool anywhere else. Not until one of the
1371          * sysctls are changed, or the surplus pages go out of use.
1372          */
1373         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1374         min_count = max(count, min_count);
1375         try_to_free_low(h, min_count, nodes_allowed);
1376         while (min_count < persistent_huge_pages(h)) {
1377                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1378                         break;
1379         }
1380         while (count < persistent_huge_pages(h)) {
1381                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1382                         break;
1383         }
1384 out:
1385         ret = persistent_huge_pages(h);
1386         spin_unlock(&hugetlb_lock);
1387         return ret;
1388 }
1389
1390 #define HSTATE_ATTR_RO(_name) \
1391         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1392
1393 #define HSTATE_ATTR(_name) \
1394         static struct kobj_attribute _name##_attr = \
1395                 __ATTR(_name, 0644, _name##_show, _name##_store)
1396
1397 static struct kobject *hugepages_kobj;
1398 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1399
1400 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1401
1402 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1403 {
1404         int i;
1405
1406         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1407                 if (hstate_kobjs[i] == kobj) {
1408                         if (nidp)
1409                                 *nidp = NUMA_NO_NODE;
1410                         return &hstates[i];
1411                 }
1412
1413         return kobj_to_node_hstate(kobj, nidp);
1414 }
1415
1416 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1417                                         struct kobj_attribute *attr, char *buf)
1418 {
1419         struct hstate *h;
1420         unsigned long nr_huge_pages;
1421         int nid;
1422
1423         h = kobj_to_hstate(kobj, &nid);
1424         if (nid == NUMA_NO_NODE)
1425                 nr_huge_pages = h->nr_huge_pages;
1426         else
1427                 nr_huge_pages = h->nr_huge_pages_node[nid];
1428
1429         return sprintf(buf, "%lu\n", nr_huge_pages);
1430 }
1431 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1432                         struct kobject *kobj, struct kobj_attribute *attr,
1433                         const char *buf, size_t len)
1434 {
1435         int err;
1436         int nid;
1437         unsigned long count;
1438         struct hstate *h;
1439         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1440
1441         err = strict_strtoul(buf, 10, &count);
1442         if (err)
1443                 return 0;
1444
1445         h = kobj_to_hstate(kobj, &nid);
1446         if (nid == NUMA_NO_NODE) {
1447                 /*
1448                  * global hstate attribute
1449                  */
1450                 if (!(obey_mempolicy &&
1451                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1452                         NODEMASK_FREE(nodes_allowed);
1453                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1454                 }
1455         } else if (nodes_allowed) {
1456                 /*
1457                  * per node hstate attribute: adjust count to global,
1458                  * but restrict alloc/free to the specified node.
1459                  */
1460                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1461                 init_nodemask_of_node(nodes_allowed, nid);
1462         } else
1463                 nodes_allowed = &node_states[N_HIGH_MEMORY];
1464
1465         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1466
1467         if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1468                 NODEMASK_FREE(nodes_allowed);
1469
1470         return len;
1471 }
1472
1473 static ssize_t nr_hugepages_show(struct kobject *kobj,
1474                                        struct kobj_attribute *attr, char *buf)
1475 {
1476         return nr_hugepages_show_common(kobj, attr, buf);
1477 }
1478
1479 static ssize_t nr_hugepages_store(struct kobject *kobj,
1480                struct kobj_attribute *attr, const char *buf, size_t len)
1481 {
1482         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1483 }
1484 HSTATE_ATTR(nr_hugepages);
1485
1486 #ifdef CONFIG_NUMA
1487
1488 /*
1489  * hstate attribute for optionally mempolicy-based constraint on persistent
1490  * huge page alloc/free.
1491  */
1492 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1493                                        struct kobj_attribute *attr, char *buf)
1494 {
1495         return nr_hugepages_show_common(kobj, attr, buf);
1496 }
1497
1498 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1499                struct kobj_attribute *attr, const char *buf, size_t len)
1500 {
1501         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1502 }
1503 HSTATE_ATTR(nr_hugepages_mempolicy);
1504 #endif
1505
1506
1507 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1508                                         struct kobj_attribute *attr, char *buf)
1509 {
1510         struct hstate *h = kobj_to_hstate(kobj, NULL);
1511         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1512 }
1513 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1514                 struct kobj_attribute *attr, const char *buf, size_t count)
1515 {
1516         int err;
1517         unsigned long input;
1518         struct hstate *h = kobj_to_hstate(kobj, NULL);
1519
1520         err = strict_strtoul(buf, 10, &input);
1521         if (err)
1522                 return 0;
1523
1524         spin_lock(&hugetlb_lock);
1525         h->nr_overcommit_huge_pages = input;
1526         spin_unlock(&hugetlb_lock);
1527
1528         return count;
1529 }
1530 HSTATE_ATTR(nr_overcommit_hugepages);
1531
1532 static ssize_t free_hugepages_show(struct kobject *kobj,
1533                                         struct kobj_attribute *attr, char *buf)
1534 {
1535         struct hstate *h;
1536         unsigned long free_huge_pages;
1537         int nid;
1538
1539         h = kobj_to_hstate(kobj, &nid);
1540         if (nid == NUMA_NO_NODE)
1541                 free_huge_pages = h->free_huge_pages;
1542         else
1543                 free_huge_pages = h->free_huge_pages_node[nid];
1544
1545         return sprintf(buf, "%lu\n", free_huge_pages);
1546 }
1547 HSTATE_ATTR_RO(free_hugepages);
1548
1549 static ssize_t resv_hugepages_show(struct kobject *kobj,
1550                                         struct kobj_attribute *attr, char *buf)
1551 {
1552         struct hstate *h = kobj_to_hstate(kobj, NULL);
1553         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1554 }
1555 HSTATE_ATTR_RO(resv_hugepages);
1556
1557 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1558                                         struct kobj_attribute *attr, char *buf)
1559 {
1560         struct hstate *h;
1561         unsigned long surplus_huge_pages;
1562         int nid;
1563
1564         h = kobj_to_hstate(kobj, &nid);
1565         if (nid == NUMA_NO_NODE)
1566                 surplus_huge_pages = h->surplus_huge_pages;
1567         else
1568                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1569
1570         return sprintf(buf, "%lu\n", surplus_huge_pages);
1571 }
1572 HSTATE_ATTR_RO(surplus_hugepages);
1573
1574 static struct attribute *hstate_attrs[] = {
1575         &nr_hugepages_attr.attr,
1576         &nr_overcommit_hugepages_attr.attr,
1577         &free_hugepages_attr.attr,
1578         &resv_hugepages_attr.attr,
1579         &surplus_hugepages_attr.attr,
1580 #ifdef CONFIG_NUMA
1581         &nr_hugepages_mempolicy_attr.attr,
1582 #endif
1583         NULL,
1584 };
1585
1586 static struct attribute_group hstate_attr_group = {
1587         .attrs = hstate_attrs,
1588 };
1589
1590 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1591                                     struct kobject **hstate_kobjs,
1592                                     struct attribute_group *hstate_attr_group)
1593 {
1594         int retval;
1595         int hi = h - hstates;
1596
1597         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1598         if (!hstate_kobjs[hi])
1599                 return -ENOMEM;
1600
1601         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1602         if (retval)
1603                 kobject_put(hstate_kobjs[hi]);
1604
1605         return retval;
1606 }
1607
1608 static void __init hugetlb_sysfs_init(void)
1609 {
1610         struct hstate *h;
1611         int err;
1612
1613         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1614         if (!hugepages_kobj)
1615                 return;
1616
1617         for_each_hstate(h) {
1618                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1619                                          hstate_kobjs, &hstate_attr_group);
1620                 if (err)
1621                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1622                                                                 h->name);
1623         }
1624 }
1625
1626 #ifdef CONFIG_NUMA
1627
1628 /*
1629  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1630  * with node sysdevs in node_devices[] using a parallel array.  The array
1631  * index of a node sysdev or _hstate == node id.
1632  * This is here to avoid any static dependency of the node sysdev driver, in
1633  * the base kernel, on the hugetlb module.
1634  */
1635 struct node_hstate {
1636         struct kobject          *hugepages_kobj;
1637         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1638 };
1639 struct node_hstate node_hstates[MAX_NUMNODES];
1640
1641 /*
1642  * A subset of global hstate attributes for node sysdevs
1643  */
1644 static struct attribute *per_node_hstate_attrs[] = {
1645         &nr_hugepages_attr.attr,
1646         &free_hugepages_attr.attr,
1647         &surplus_hugepages_attr.attr,
1648         NULL,
1649 };
1650
1651 static struct attribute_group per_node_hstate_attr_group = {
1652         .attrs = per_node_hstate_attrs,
1653 };
1654
1655 /*
1656  * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1657  * Returns node id via non-NULL nidp.
1658  */
1659 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1660 {
1661         int nid;
1662
1663         for (nid = 0; nid < nr_node_ids; nid++) {
1664                 struct node_hstate *nhs = &node_hstates[nid];
1665                 int i;
1666                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1667                         if (nhs->hstate_kobjs[i] == kobj) {
1668                                 if (nidp)
1669                                         *nidp = nid;
1670                                 return &hstates[i];
1671                         }
1672         }
1673
1674         BUG();
1675         return NULL;
1676 }
1677
1678 /*
1679  * Unregister hstate attributes from a single node sysdev.
1680  * No-op if no hstate attributes attached.
1681  */
1682 void hugetlb_unregister_node(struct node *node)
1683 {
1684         struct hstate *h;
1685         struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1686
1687         if (!nhs->hugepages_kobj)
1688                 return;         /* no hstate attributes */
1689
1690         for_each_hstate(h)
1691                 if (nhs->hstate_kobjs[h - hstates]) {
1692                         kobject_put(nhs->hstate_kobjs[h - hstates]);
1693                         nhs->hstate_kobjs[h - hstates] = NULL;
1694                 }
1695
1696         kobject_put(nhs->hugepages_kobj);
1697         nhs->hugepages_kobj = NULL;
1698 }
1699
1700 /*
1701  * hugetlb module exit:  unregister hstate attributes from node sysdevs
1702  * that have them.
1703  */
1704 static void hugetlb_unregister_all_nodes(void)
1705 {
1706         int nid;
1707
1708         /*
1709          * disable node sysdev registrations.
1710          */
1711         register_hugetlbfs_with_node(NULL, NULL);
1712
1713         /*
1714          * remove hstate attributes from any nodes that have them.
1715          */
1716         for (nid = 0; nid < nr_node_ids; nid++)
1717                 hugetlb_unregister_node(&node_devices[nid]);
1718 }
1719
1720 /*
1721  * Register hstate attributes for a single node sysdev.
1722  * No-op if attributes already registered.
1723  */
1724 void hugetlb_register_node(struct node *node)
1725 {
1726         struct hstate *h;
1727         struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1728         int err;
1729
1730         if (nhs->hugepages_kobj)
1731                 return;         /* already allocated */
1732
1733         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1734                                                         &node->sysdev.kobj);
1735         if (!nhs->hugepages_kobj)
1736                 return;
1737
1738         for_each_hstate(h) {
1739                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1740                                                 nhs->hstate_kobjs,
1741                                                 &per_node_hstate_attr_group);
1742                 if (err) {
1743                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1744                                         " for node %d\n",
1745                                                 h->name, node->sysdev.id);
1746                         hugetlb_unregister_node(node);
1747                         break;
1748                 }
1749         }
1750 }
1751
1752 /*
1753  * hugetlb init time:  register hstate attributes for all registered node
1754  * sysdevs of nodes that have memory.  All on-line nodes should have
1755  * registered their associated sysdev by this time.
1756  */
1757 static void hugetlb_register_all_nodes(void)
1758 {
1759         int nid;
1760
1761         for_each_node_state(nid, N_HIGH_MEMORY) {
1762                 struct node *node = &node_devices[nid];
1763                 if (node->sysdev.id == nid)
1764                         hugetlb_register_node(node);
1765         }
1766
1767         /*
1768          * Let the node sysdev driver know we're here so it can
1769          * [un]register hstate attributes on node hotplug.
1770          */
1771         register_hugetlbfs_with_node(hugetlb_register_node,
1772                                      hugetlb_unregister_node);
1773 }
1774 #else   /* !CONFIG_NUMA */
1775
1776 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1777 {
1778         BUG();
1779         if (nidp)
1780                 *nidp = -1;
1781         return NULL;
1782 }
1783
1784 static void hugetlb_unregister_all_nodes(void) { }
1785
1786 static void hugetlb_register_all_nodes(void) { }
1787
1788 #endif
1789
1790 static void __exit hugetlb_exit(void)
1791 {
1792         struct hstate *h;
1793
1794         hugetlb_unregister_all_nodes();
1795
1796         for_each_hstate(h) {
1797                 kobject_put(hstate_kobjs[h - hstates]);
1798         }
1799
1800         kobject_put(hugepages_kobj);
1801 }
1802 module_exit(hugetlb_exit);
1803
1804 static int __init hugetlb_init(void)
1805 {
1806         /* Some platform decide whether they support huge pages at boot
1807          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1808          * there is no such support
1809          */
1810         if (HPAGE_SHIFT == 0)
1811                 return 0;
1812
1813         if (!size_to_hstate(default_hstate_size)) {
1814                 default_hstate_size = HPAGE_SIZE;
1815                 if (!size_to_hstate(default_hstate_size))
1816                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1817         }
1818         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1819         if (default_hstate_max_huge_pages)
1820                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1821
1822         hugetlb_init_hstates();
1823
1824         gather_bootmem_prealloc();
1825
1826         report_hugepages();
1827
1828         hugetlb_sysfs_init();
1829
1830         hugetlb_register_all_nodes();
1831
1832         return 0;
1833 }
1834 module_init(hugetlb_init);
1835
1836 /* Should be called on processing a hugepagesz=... option */
1837 void __init hugetlb_add_hstate(unsigned order)
1838 {
1839         struct hstate *h;
1840         unsigned long i;
1841
1842         if (size_to_hstate(PAGE_SIZE << order)) {
1843                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1844                 return;
1845         }
1846         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1847         BUG_ON(order == 0);
1848         h = &hstates[max_hstate++];
1849         h->order = order;
1850         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1851         h->nr_huge_pages = 0;
1852         h->free_huge_pages = 0;
1853         for (i = 0; i < MAX_NUMNODES; ++i)
1854                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1855         h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1856         h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1857         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1858                                         huge_page_size(h)/1024);
1859
1860         parsed_hstate = h;
1861 }
1862
1863 static int __init hugetlb_nrpages_setup(char *s)
1864 {
1865         unsigned long *mhp;
1866         static unsigned long *last_mhp;
1867
1868         /*
1869          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1870          * so this hugepages= parameter goes to the "default hstate".
1871          */
1872         if (!max_hstate)
1873                 mhp = &default_hstate_max_huge_pages;
1874         else
1875                 mhp = &parsed_hstate->max_huge_pages;
1876
1877         if (mhp == last_mhp) {
1878                 printk(KERN_WARNING "hugepages= specified twice without "
1879                         "interleaving hugepagesz=, ignoring\n");
1880                 return 1;
1881         }
1882
1883         if (sscanf(s, "%lu", mhp) <= 0)
1884                 *mhp = 0;
1885
1886         /*
1887          * Global state is always initialized later in hugetlb_init.
1888          * But we need to allocate >= MAX_ORDER hstates here early to still
1889          * use the bootmem allocator.
1890          */
1891         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1892                 hugetlb_hstate_alloc_pages(parsed_hstate);
1893
1894         last_mhp = mhp;
1895
1896         return 1;
1897 }
1898 __setup("hugepages=", hugetlb_nrpages_setup);
1899
1900 static int __init hugetlb_default_setup(char *s)
1901 {
1902         default_hstate_size = memparse(s, &s);
1903         return 1;
1904 }
1905 __setup("default_hugepagesz=", hugetlb_default_setup);
1906
1907 static unsigned int cpuset_mems_nr(unsigned int *array)
1908 {
1909         int node;
1910         unsigned int nr = 0;
1911
1912         for_each_node_mask(node, cpuset_current_mems_allowed)
1913                 nr += array[node];
1914
1915         return nr;
1916 }
1917
1918 #ifdef CONFIG_SYSCTL
1919 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1920                          struct ctl_table *table, int write,
1921                          void __user *buffer, size_t *length, loff_t *ppos)
1922 {
1923         struct hstate *h = &default_hstate;
1924         unsigned long tmp;
1925
1926         if (!write)
1927                 tmp = h->max_huge_pages;
1928
1929         table->data = &tmp;
1930         table->maxlen = sizeof(unsigned long);
1931         proc_doulongvec_minmax(table, write, buffer, length, ppos);
1932
1933         if (write) {
1934                 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1935                                                 GFP_KERNEL | __GFP_NORETRY);
1936                 if (!(obey_mempolicy &&
1937                                init_nodemask_of_mempolicy(nodes_allowed))) {
1938                         NODEMASK_FREE(nodes_allowed);
1939                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1940                 }
1941                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1942
1943                 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1944                         NODEMASK_FREE(nodes_allowed);
1945         }
1946
1947         return 0;
1948 }
1949
1950 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1951                           void __user *buffer, size_t *length, loff_t *ppos)
1952 {
1953
1954         return hugetlb_sysctl_handler_common(false, table, write,
1955                                                         buffer, length, ppos);
1956 }
1957
1958 #ifdef CONFIG_NUMA
1959 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1960                           void __user *buffer, size_t *length, loff_t *ppos)
1961 {
1962         return hugetlb_sysctl_handler_common(true, table, write,
1963                                                         buffer, length, ppos);
1964 }
1965 #endif /* CONFIG_NUMA */
1966
1967 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1968                         void __user *buffer,
1969                         size_t *length, loff_t *ppos)
1970 {
1971         proc_dointvec(table, write, buffer, length, ppos);
1972         if (hugepages_treat_as_movable)
1973                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1974         else
1975                 htlb_alloc_mask = GFP_HIGHUSER;
1976         return 0;
1977 }
1978
1979 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1980                         void __user *buffer,
1981                         size_t *length, loff_t *ppos)
1982 {
1983         struct hstate *h = &default_hstate;
1984         unsigned long tmp;
1985
1986         if (!write)
1987                 tmp = h->nr_overcommit_huge_pages;
1988
1989         table->data = &tmp;
1990         table->maxlen = sizeof(unsigned long);
1991         proc_doulongvec_minmax(table, write, buffer, length, ppos);
1992
1993         if (write) {
1994                 spin_lock(&hugetlb_lock);
1995                 h->nr_overcommit_huge_pages = tmp;
1996                 spin_unlock(&hugetlb_lock);
1997         }
1998
1999         return 0;
2000 }
2001
2002 #endif /* CONFIG_SYSCTL */
2003
2004 void hugetlb_report_meminfo(struct seq_file *m)
2005 {
2006         struct hstate *h = &default_hstate;
2007         seq_printf(m,
2008                         "HugePages_Total:   %5lu\n"
2009                         "HugePages_Free:    %5lu\n"
2010                         "HugePages_Rsvd:    %5lu\n"
2011                         "HugePages_Surp:    %5lu\n"
2012                         "Hugepagesize:   %8lu kB\n",
2013                         h->nr_huge_pages,
2014                         h->free_huge_pages,
2015                         h->resv_huge_pages,
2016                         h->surplus_huge_pages,
2017                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2018 }
2019
2020 int hugetlb_report_node_meminfo(int nid, char *buf)
2021 {
2022         struct hstate *h = &default_hstate;
2023         return sprintf(buf,
2024                 "Node %d HugePages_Total: %5u\n"
2025                 "Node %d HugePages_Free:  %5u\n"
2026                 "Node %d HugePages_Surp:  %5u\n",
2027                 nid, h->nr_huge_pages_node[nid],
2028                 nid, h->free_huge_pages_node[nid],
2029                 nid, h->surplus_huge_pages_node[nid]);
2030 }
2031
2032 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2033 unsigned long hugetlb_total_pages(void)
2034 {
2035         struct hstate *h = &default_hstate;
2036         return h->nr_huge_pages * pages_per_huge_page(h);
2037 }
2038
2039 static int hugetlb_acct_memory(struct hstate *h, long delta)
2040 {
2041         int ret = -ENOMEM;
2042
2043         spin_lock(&hugetlb_lock);
2044         /*
2045          * When cpuset is configured, it breaks the strict hugetlb page
2046          * reservation as the accounting is done on a global variable. Such
2047          * reservation is completely rubbish in the presence of cpuset because
2048          * the reservation is not checked against page availability for the
2049          * current cpuset. Application can still potentially OOM'ed by kernel
2050          * with lack of free htlb page in cpuset that the task is in.
2051          * Attempt to enforce strict accounting with cpuset is almost
2052          * impossible (or too ugly) because cpuset is too fluid that
2053          * task or memory node can be dynamically moved between cpusets.
2054          *
2055          * The change of semantics for shared hugetlb mapping with cpuset is
2056          * undesirable. However, in order to preserve some of the semantics,
2057          * we fall back to check against current free page availability as
2058          * a best attempt and hopefully to minimize the impact of changing
2059          * semantics that cpuset has.
2060          */
2061         if (delta > 0) {
2062                 if (gather_surplus_pages(h, delta) < 0)
2063                         goto out;
2064
2065                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2066                         return_unused_surplus_pages(h, delta);
2067                         goto out;
2068                 }
2069         }
2070
2071         ret = 0;
2072         if (delta < 0)
2073                 return_unused_surplus_pages(h, (unsigned long) -delta);
2074
2075 out:
2076         spin_unlock(&hugetlb_lock);
2077         return ret;
2078 }
2079
2080 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2081 {
2082         struct resv_map *reservations = vma_resv_map(vma);
2083
2084         /*
2085          * This new VMA should share its siblings reservation map if present.
2086          * The VMA will only ever have a valid reservation map pointer where
2087          * it is being copied for another still existing VMA.  As that VMA
2088          * has a reference to the reservation map it cannot dissappear until
2089          * after this open call completes.  It is therefore safe to take a
2090          * new reference here without additional locking.
2091          */
2092         if (reservations)
2093                 kref_get(&reservations->refs);
2094 }
2095
2096 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2097 {
2098         struct hstate *h = hstate_vma(vma);
2099         struct resv_map *reservations = vma_resv_map(vma);
2100         unsigned long reserve;
2101         unsigned long start;
2102         unsigned long end;
2103
2104         if (reservations) {
2105                 start = vma_hugecache_offset(h, vma, vma->vm_start);
2106                 end = vma_hugecache_offset(h, vma, vma->vm_end);
2107
2108                 reserve = (end - start) -
2109                         region_count(&reservations->regions, start, end);
2110
2111                 kref_put(&reservations->refs, resv_map_release);
2112
2113                 if (reserve) {
2114                         hugetlb_acct_memory(h, -reserve);
2115                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2116                 }
2117         }
2118 }
2119
2120 /*
2121  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2122  * handle_mm_fault() to try to instantiate regular-sized pages in the
2123  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2124  * this far.
2125  */
2126 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2127 {
2128         BUG();
2129         return 0;
2130 }
2131
2132 const struct vm_operations_struct hugetlb_vm_ops = {
2133         .fault = hugetlb_vm_op_fault,
2134         .open = hugetlb_vm_op_open,
2135         .close = hugetlb_vm_op_close,
2136 };
2137
2138 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2139                                 int writable)
2140 {
2141         pte_t entry;
2142
2143         if (writable) {
2144                 entry =
2145                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2146         } else {
2147                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2148         }
2149         entry = pte_mkyoung(entry);
2150         entry = pte_mkhuge(entry);
2151
2152         return entry;
2153 }
2154
2155 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2156                                    unsigned long address, pte_t *ptep)
2157 {
2158         pte_t entry;
2159
2160         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2161         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
2162                 update_mmu_cache(vma, address, ptep);
2163         }
2164 }
2165
2166
2167 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2168                             struct vm_area_struct *vma)
2169 {
2170         pte_t *src_pte, *dst_pte, entry;
2171         struct page *ptepage;
2172         unsigned long addr;
2173         int cow;
2174         struct hstate *h = hstate_vma(vma);
2175         unsigned long sz = huge_page_size(h);
2176
2177         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2178
2179         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2180                 src_pte = huge_pte_offset(src, addr);
2181                 if (!src_pte)
2182                         continue;
2183                 dst_pte = huge_pte_alloc(dst, addr, sz);
2184                 if (!dst_pte)
2185                         goto nomem;
2186
2187                 /* If the pagetables are shared don't copy or take references */
2188                 if (dst_pte == src_pte)
2189                         continue;
2190
2191                 spin_lock(&dst->page_table_lock);
2192                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2193                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2194                         if (cow)
2195                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2196                         entry = huge_ptep_get(src_pte);
2197                         ptepage = pte_page(entry);
2198                         get_page(ptepage);
2199                         page_dup_rmap(ptepage);
2200                         set_huge_pte_at(dst, addr, dst_pte, entry);
2201                 }
2202                 spin_unlock(&src->page_table_lock);
2203                 spin_unlock(&dst->page_table_lock);
2204         }
2205         return 0;
2206
2207 nomem:
2208         return -ENOMEM;
2209 }
2210
2211 static int is_hugetlb_entry_migration(pte_t pte)
2212 {
2213         swp_entry_t swp;
2214
2215         if (huge_pte_none(pte) || pte_present(pte))
2216                 return 0;
2217         swp = pte_to_swp_entry(pte);
2218         if (non_swap_entry(swp) && is_migration_entry(swp)) {
2219                 return 1;
2220         } else
2221                 return 0;
2222 }
2223
2224 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2225 {
2226         swp_entry_t swp;
2227
2228         if (huge_pte_none(pte) || pte_present(pte))
2229                 return 0;
2230         swp = pte_to_swp_entry(pte);
2231         if (non_swap_entry(swp) && is_hwpoison_entry(swp)) {
2232                 return 1;
2233         } else
2234                 return 0;
2235 }
2236
2237 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2238                             unsigned long end, struct page *ref_page)
2239 {
2240         struct mm_struct *mm = vma->vm_mm;
2241         unsigned long address;
2242         pte_t *ptep;
2243         pte_t pte;
2244         struct page *page;
2245         struct page *tmp;
2246         struct hstate *h = hstate_vma(vma);
2247         unsigned long sz = huge_page_size(h);
2248
2249         /*
2250          * A page gathering list, protected by per file i_mmap_lock. The
2251          * lock is used to avoid list corruption from multiple unmapping
2252          * of the same page since we are using page->lru.
2253          */
2254         LIST_HEAD(page_list);
2255
2256         WARN_ON(!is_vm_hugetlb_page(vma));
2257         BUG_ON(start & ~huge_page_mask(h));
2258         BUG_ON(end & ~huge_page_mask(h));
2259
2260         mmu_notifier_invalidate_range_start(mm, start, end);
2261         spin_lock(&mm->page_table_lock);
2262         for (address = start; address < end; address += sz) {
2263                 ptep = huge_pte_offset(mm, address);
2264                 if (!ptep)
2265                         continue;
2266
2267                 if (huge_pmd_unshare(mm, &address, ptep))
2268                         continue;
2269
2270                 /*
2271                  * If a reference page is supplied, it is because a specific
2272                  * page is being unmapped, not a range. Ensure the page we
2273                  * are about to unmap is the actual page of interest.
2274                  */
2275                 if (ref_page) {
2276                         pte = huge_ptep_get(ptep);
2277                         if (huge_pte_none(pte))
2278                                 continue;
2279                         page = pte_page(pte);
2280                         if (page != ref_page)
2281                                 continue;
2282
2283                         /*
2284                          * Mark the VMA as having unmapped its page so that
2285                          * future faults in this VMA will fail rather than
2286                          * looking like data was lost
2287                          */
2288                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2289                 }
2290
2291                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2292                 if (huge_pte_none(pte))
2293                         continue;
2294
2295                 /*
2296                  * HWPoisoned hugepage is already unmapped and dropped reference
2297                  */
2298                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2299                         continue;
2300
2301                 page = pte_page(pte);
2302                 if (pte_dirty(pte))
2303                         set_page_dirty(page);
2304                 list_add(&page->lru, &page_list);
2305         }
2306         spin_unlock(&mm->page_table_lock);
2307         flush_tlb_range(vma, start, end);
2308         mmu_notifier_invalidate_range_end(mm, start, end);
2309         list_for_each_entry_safe(page, tmp, &page_list, lru) {
2310                 page_remove_rmap(page);
2311                 list_del(&page->lru);
2312                 put_page(page);
2313         }
2314 }
2315
2316 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2317                           unsigned long end, struct page *ref_page)
2318 {
2319         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2320         __unmap_hugepage_range(vma, start, end, ref_page);
2321         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2322 }
2323
2324 /*
2325  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2326  * mappping it owns the reserve page for. The intention is to unmap the page
2327  * from other VMAs and let the children be SIGKILLed if they are faulting the
2328  * same region.
2329  */
2330 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2331                                 struct page *page, unsigned long address)
2332 {
2333         struct hstate *h = hstate_vma(vma);
2334         struct vm_area_struct *iter_vma;
2335         struct address_space *mapping;
2336         struct prio_tree_iter iter;
2337         pgoff_t pgoff;
2338
2339         /*
2340          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2341          * from page cache lookup which is in HPAGE_SIZE units.
2342          */
2343         address = address & huge_page_mask(h);
2344         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2345                 + (vma->vm_pgoff >> PAGE_SHIFT);
2346         mapping = (struct address_space *)page_private(page);
2347
2348         /*
2349          * Take the mapping lock for the duration of the table walk. As
2350          * this mapping should be shared between all the VMAs,
2351          * __unmap_hugepage_range() is called as the lock is already held
2352          */
2353         spin_lock(&mapping->i_mmap_lock);
2354         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2355                 /* Do not unmap the current VMA */
2356                 if (iter_vma == vma)
2357                         continue;
2358
2359                 /*
2360                  * Unmap the page from other VMAs without their own reserves.
2361                  * They get marked to be SIGKILLed if they fault in these
2362                  * areas. This is because a future no-page fault on this VMA
2363                  * could insert a zeroed page instead of the data existing
2364                  * from the time of fork. This would look like data corruption
2365                  */
2366                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2367                         __unmap_hugepage_range(iter_vma,
2368                                 address, address + huge_page_size(h),
2369                                 page);
2370         }
2371         spin_unlock(&mapping->i_mmap_lock);
2372
2373         return 1;
2374 }
2375
2376 /*
2377  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2378  */
2379 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2380                         unsigned long address, pte_t *ptep, pte_t pte,
2381                         struct page *pagecache_page)
2382 {
2383         struct hstate *h = hstate_vma(vma);
2384         struct page *old_page, *new_page;
2385         int avoidcopy;
2386         int outside_reserve = 0;
2387
2388         old_page = pte_page(pte);
2389
2390 retry_avoidcopy:
2391         /* If no-one else is actually using this page, avoid the copy
2392          * and just make the page writable */
2393         avoidcopy = (page_mapcount(old_page) == 1);
2394         if (avoidcopy) {
2395                 if (PageAnon(old_page))
2396                         page_move_anon_rmap(old_page, vma, address);
2397                 set_huge_ptep_writable(vma, address, ptep);
2398                 return 0;
2399         }
2400
2401         /*
2402          * If the process that created a MAP_PRIVATE mapping is about to
2403          * perform a COW due to a shared page count, attempt to satisfy
2404          * the allocation without using the existing reserves. The pagecache
2405          * page is used to determine if the reserve at this address was
2406          * consumed or not. If reserves were used, a partial faulted mapping
2407          * at the time of fork() could consume its reserves on COW instead
2408          * of the full address range.
2409          */
2410         if (!(vma->vm_flags & VM_MAYSHARE) &&
2411                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2412                         old_page != pagecache_page)
2413                 outside_reserve = 1;
2414
2415         page_cache_get(old_page);
2416
2417         /* Drop page_table_lock as buddy allocator may be called */
2418         spin_unlock(&mm->page_table_lock);
2419         new_page = alloc_huge_page(vma, address, outside_reserve);
2420
2421         if (IS_ERR(new_page)) {
2422                 page_cache_release(old_page);
2423
2424                 /*
2425                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2426                  * it is due to references held by a child and an insufficient
2427                  * huge page pool. To guarantee the original mappers
2428                  * reliability, unmap the page from child processes. The child
2429                  * may get SIGKILLed if it later faults.
2430                  */
2431                 if (outside_reserve) {
2432                         BUG_ON(huge_pte_none(pte));
2433                         if (unmap_ref_private(mm, vma, old_page, address)) {
2434                                 BUG_ON(page_count(old_page) != 1);
2435                                 BUG_ON(huge_pte_none(pte));
2436                                 spin_lock(&mm->page_table_lock);
2437                                 goto retry_avoidcopy;
2438                         }
2439                         WARN_ON_ONCE(1);
2440                 }
2441
2442                 /* Caller expects lock to be held */
2443                 spin_lock(&mm->page_table_lock);
2444                 return -PTR_ERR(new_page);
2445         }
2446
2447         /*
2448          * When the original hugepage is shared one, it does not have
2449          * anon_vma prepared.
2450          */
2451         if (unlikely(anon_vma_prepare(vma))) {
2452                 /* Caller expects lock to be held */
2453                 spin_lock(&mm->page_table_lock);
2454                 return VM_FAULT_OOM;
2455         }
2456
2457         copy_user_huge_page(new_page, old_page, address, vma);
2458         __SetPageUptodate(new_page);
2459
2460         /*
2461          * Retake the page_table_lock to check for racing updates
2462          * before the page tables are altered
2463          */
2464         spin_lock(&mm->page_table_lock);
2465         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2466         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2467                 /* Break COW */
2468                 mmu_notifier_invalidate_range_start(mm,
2469                         address & huge_page_mask(h),
2470                         (address & huge_page_mask(h)) + huge_page_size(h));
2471                 huge_ptep_clear_flush(vma, address, ptep);
2472                 set_huge_pte_at(mm, address, ptep,
2473                                 make_huge_pte(vma, new_page, 1));
2474                 page_remove_rmap(old_page);
2475                 hugepage_add_new_anon_rmap(new_page, vma, address);
2476                 /* Make the old page be freed below */
2477                 new_page = old_page;
2478                 mmu_notifier_invalidate_range_end(mm,
2479                         address & huge_page_mask(h),
2480                         (address & huge_page_mask(h)) + huge_page_size(h));
2481         }
2482         page_cache_release(new_page);
2483         page_cache_release(old_page);
2484         return 0;
2485 }
2486
2487 /* Return the pagecache page at a given address within a VMA */
2488 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2489                         struct vm_area_struct *vma, unsigned long address)
2490 {
2491         struct address_space *mapping;
2492         pgoff_t idx;
2493
2494         mapping = vma->vm_file->f_mapping;
2495         idx = vma_hugecache_offset(h, vma, address);
2496
2497         return find_lock_page(mapping, idx);
2498 }
2499
2500 /*
2501  * Return whether there is a pagecache page to back given address within VMA.
2502  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2503  */
2504 static bool hugetlbfs_pagecache_present(struct hstate *h,
2505                         struct vm_area_struct *vma, unsigned long address)
2506 {
2507         struct address_space *mapping;
2508         pgoff_t idx;
2509         struct page *page;
2510
2511         mapping = vma->vm_file->f_mapping;
2512         idx = vma_hugecache_offset(h, vma, address);
2513
2514         page = find_get_page(mapping, idx);
2515         if (page)
2516                 put_page(page);
2517         return page != NULL;
2518 }
2519
2520 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2521                         unsigned long address, pte_t *ptep, unsigned int flags)
2522 {
2523         struct hstate *h = hstate_vma(vma);
2524         int ret = VM_FAULT_SIGBUS;
2525         pgoff_t idx;
2526         unsigned long size;
2527         struct page *page;
2528         struct address_space *mapping;
2529         pte_t new_pte;
2530
2531         /*
2532          * Currently, we are forced to kill the process in the event the
2533          * original mapper has unmapped pages from the child due to a failed
2534          * COW. Warn that such a situation has occured as it may not be obvious
2535          */
2536         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2537                 printk(KERN_WARNING
2538                         "PID %d killed due to inadequate hugepage pool\n",
2539                         current->pid);
2540                 return ret;
2541         }
2542
2543         mapping = vma->vm_file->f_mapping;
2544         idx = vma_hugecache_offset(h, vma, address);
2545
2546         /*
2547          * Use page lock to guard against racing truncation
2548          * before we get page_table_lock.
2549          */
2550 retry:
2551         page = find_lock_page(mapping, idx);
2552         if (!page) {
2553                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2554                 if (idx >= size)
2555                         goto out;
2556                 page = alloc_huge_page(vma, address, 0);
2557                 if (IS_ERR(page)) {
2558                         ret = -PTR_ERR(page);
2559                         goto out;
2560                 }
2561                 clear_huge_page(page, address, huge_page_size(h));
2562                 __SetPageUptodate(page);
2563
2564                 if (vma->vm_flags & VM_MAYSHARE) {
2565                         int err;
2566                         struct inode *inode = mapping->host;
2567
2568                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2569                         if (err) {
2570                                 put_page(page);
2571                                 if (err == -EEXIST)
2572                                         goto retry;
2573                                 goto out;
2574                         }
2575
2576                         spin_lock(&inode->i_lock);
2577                         inode->i_blocks += blocks_per_huge_page(h);
2578                         spin_unlock(&inode->i_lock);
2579                         page_dup_rmap(page);
2580                 } else {
2581                         lock_page(page);
2582                         if (unlikely(anon_vma_prepare(vma))) {
2583                                 ret = VM_FAULT_OOM;
2584                                 goto backout_unlocked;
2585                         }
2586                         hugepage_add_new_anon_rmap(page, vma, address);
2587                 }
2588         } else {
2589                 /*
2590                  * If memory error occurs between mmap() and fault, some process
2591                  * don't have hwpoisoned swap entry for errored virtual address.
2592                  * So we need to block hugepage fault by PG_hwpoison bit check.
2593                  */
2594                 if (unlikely(PageHWPoison(page))) {
2595                         ret = VM_FAULT_HWPOISON | 
2596                               VM_FAULT_SET_HINDEX(h - hstates);
2597                         goto backout_unlocked;
2598                 }
2599                 page_dup_rmap(page);
2600         }
2601
2602         /*
2603          * If we are going to COW a private mapping later, we examine the
2604          * pending reservations for this page now. This will ensure that
2605          * any allocations necessary to record that reservation occur outside
2606          * the spinlock.
2607          */
2608         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2609                 if (vma_needs_reservation(h, vma, address) < 0) {
2610                         ret = VM_FAULT_OOM;
2611                         goto backout_unlocked;
2612                 }
2613
2614         spin_lock(&mm->page_table_lock);
2615         size = i_size_read(mapping->host) >> huge_page_shift(h);
2616         if (idx >= size)
2617                 goto backout;
2618
2619         ret = 0;
2620         if (!huge_pte_none(huge_ptep_get(ptep)))
2621                 goto backout;
2622
2623         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2624                                 && (vma->vm_flags & VM_SHARED)));
2625         set_huge_pte_at(mm, address, ptep, new_pte);
2626
2627         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2628                 /* Optimization, do the COW without a second fault */
2629                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2630         }
2631
2632         spin_unlock(&mm->page_table_lock);
2633         unlock_page(page);
2634 out:
2635         return ret;
2636
2637 backout:
2638         spin_unlock(&mm->page_table_lock);
2639 backout_unlocked:
2640         unlock_page(page);
2641         put_page(page);
2642         goto out;
2643 }
2644
2645 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2646                         unsigned long address, unsigned int flags)
2647 {
2648         pte_t *ptep;
2649         pte_t entry;
2650         int ret;
2651         struct page *page = NULL;
2652         struct page *pagecache_page = NULL;
2653         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2654         struct hstate *h = hstate_vma(vma);
2655
2656         ptep = huge_pte_offset(mm, address);
2657         if (ptep) {
2658                 entry = huge_ptep_get(ptep);
2659                 if (unlikely(is_hugetlb_entry_migration(entry))) {
2660                         migration_entry_wait(mm, (pmd_t *)ptep, address);
2661                         return 0;
2662                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2663                         return VM_FAULT_HWPOISON_LARGE | 
2664                                VM_FAULT_SET_HINDEX(h - hstates);
2665         }
2666
2667         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2668         if (!ptep)
2669                 return VM_FAULT_OOM;
2670
2671         /*
2672          * Serialize hugepage allocation and instantiation, so that we don't
2673          * get spurious allocation failures if two CPUs race to instantiate
2674          * the same page in the page cache.
2675          */
2676         mutex_lock(&hugetlb_instantiation_mutex);
2677         entry = huge_ptep_get(ptep);
2678         if (huge_pte_none(entry)) {
2679                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2680                 goto out_mutex;
2681         }
2682
2683         ret = 0;
2684
2685         /*
2686          * If we are going to COW the mapping later, we examine the pending
2687          * reservations for this page now. This will ensure that any
2688          * allocations necessary to record that reservation occur outside the
2689          * spinlock. For private mappings, we also lookup the pagecache
2690          * page now as it is used to determine if a reservation has been
2691          * consumed.
2692          */
2693         if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2694                 if (vma_needs_reservation(h, vma, address) < 0) {
2695                         ret = VM_FAULT_OOM;
2696                         goto out_mutex;
2697                 }
2698
2699                 if (!(vma->vm_flags & VM_MAYSHARE))
2700                         pagecache_page = hugetlbfs_pagecache_page(h,
2701                                                                 vma, address);
2702         }
2703
2704         /*
2705          * hugetlb_cow() requires page locks of pte_page(entry) and
2706          * pagecache_page, so here we need take the former one
2707          * when page != pagecache_page or !pagecache_page.
2708          * Note that locking order is always pagecache_page -> page,
2709          * so no worry about deadlock.
2710          */
2711         page = pte_page(entry);
2712         if (page != pagecache_page)
2713                 lock_page(page);
2714
2715         spin_lock(&mm->page_table_lock);
2716         /* Check for a racing update before calling hugetlb_cow */
2717         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2718                 goto out_page_table_lock;
2719
2720
2721         if (flags & FAULT_FLAG_WRITE) {
2722                 if (!pte_write(entry)) {
2723                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2724                                                         pagecache_page);
2725                         goto out_page_table_lock;
2726                 }
2727                 entry = pte_mkdirty(entry);
2728         }
2729         entry = pte_mkyoung(entry);
2730         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2731                                                 flags & FAULT_FLAG_WRITE))
2732                 update_mmu_cache(vma, address, ptep);
2733
2734 out_page_table_lock:
2735         spin_unlock(&mm->page_table_lock);
2736
2737         if (pagecache_page) {
2738                 unlock_page(pagecache_page);
2739                 put_page(pagecache_page);
2740         }
2741         unlock_page(page);
2742
2743 out_mutex:
2744         mutex_unlock(&hugetlb_instantiation_mutex);
2745
2746         return ret;
2747 }
2748
2749 /* Can be overriden by architectures */
2750 __attribute__((weak)) struct page *
2751 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2752                pud_t *pud, int write)
2753 {
2754         BUG();
2755         return NULL;
2756 }
2757
2758 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2759                         struct page **pages, struct vm_area_struct **vmas,
2760                         unsigned long *position, int *length, int i,
2761                         unsigned int flags)
2762 {
2763         unsigned long pfn_offset;
2764         unsigned long vaddr = *position;
2765         int remainder = *length;
2766         struct hstate *h = hstate_vma(vma);
2767
2768         spin_lock(&mm->page_table_lock);
2769         while (vaddr < vma->vm_end && remainder) {
2770                 pte_t *pte;
2771                 int absent;
2772                 struct page *page;
2773
2774                 /*
2775                  * Some archs (sparc64, sh*) have multiple pte_ts to
2776                  * each hugepage.  We have to make sure we get the
2777                  * first, for the page indexing below to work.
2778                  */
2779                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2780                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2781
2782                 /*
2783                  * When coredumping, it suits get_dump_page if we just return
2784                  * an error where there's an empty slot with no huge pagecache
2785                  * to back it.  This way, we avoid allocating a hugepage, and
2786                  * the sparse dumpfile avoids allocating disk blocks, but its
2787                  * huge holes still show up with zeroes where they need to be.
2788                  */
2789                 if (absent && (flags & FOLL_DUMP) &&
2790                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2791                         remainder = 0;
2792                         break;
2793                 }
2794
2795                 if (absent ||
2796                     ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2797                         int ret;
2798
2799                         spin_unlock(&mm->page_table_lock);
2800                         ret = hugetlb_fault(mm, vma, vaddr,
2801                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2802                         spin_lock(&mm->page_table_lock);
2803                         if (!(ret & VM_FAULT_ERROR))
2804                                 continue;
2805
2806                         remainder = 0;
2807                         break;
2808                 }
2809
2810                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2811                 page = pte_page(huge_ptep_get(pte));
2812 same_page:
2813                 if (pages) {
2814                         pages[i] = mem_map_offset(page, pfn_offset);
2815                         get_page(pages[i]);
2816                 }
2817
2818                 if (vmas)
2819                         vmas[i] = vma;
2820
2821                 vaddr += PAGE_SIZE;
2822                 ++pfn_offset;
2823                 --remainder;
2824                 ++i;
2825                 if (vaddr < vma->vm_end && remainder &&
2826                                 pfn_offset < pages_per_huge_page(h)) {
2827                         /*
2828                          * We use pfn_offset to avoid touching the pageframes
2829                          * of this compound page.
2830                          */
2831                         goto same_page;
2832                 }
2833         }
2834         spin_unlock(&mm->page_table_lock);
2835         *length = remainder;
2836         *position = vaddr;
2837
2838         return i ? i : -EFAULT;
2839 }
2840
2841 void hugetlb_change_protection(struct vm_area_struct *vma,
2842                 unsigned long address, unsigned long end, pgprot_t newprot)
2843 {
2844         struct mm_struct *mm = vma->vm_mm;
2845         unsigned long start = address;
2846         pte_t *ptep;
2847         pte_t pte;
2848         struct hstate *h = hstate_vma(vma);
2849
2850         BUG_ON(address >= end);
2851         flush_cache_range(vma, address, end);
2852
2853         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2854         spin_lock(&mm->page_table_lock);
2855         for (; address < end; address += huge_page_size(h)) {
2856                 ptep = huge_pte_offset(mm, address);
2857                 if (!ptep)
2858                         continue;
2859                 if (huge_pmd_unshare(mm, &address, ptep))
2860                         continue;
2861                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2862                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2863                         pte = pte_mkhuge(pte_modify(pte, newprot));
2864                         set_huge_pte_at(mm, address, ptep, pte);
2865                 }
2866         }
2867         spin_unlock(&mm->page_table_lock);
2868         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2869
2870         flush_tlb_range(vma, start, end);
2871 }
2872
2873 int hugetlb_reserve_pages(struct inode *inode,
2874                                         long from, long to,
2875                                         struct vm_area_struct *vma,
2876                                         int acctflag)
2877 {
2878         long ret, chg;
2879         struct hstate *h = hstate_inode(inode);
2880
2881         /*
2882          * Only apply hugepage reservation if asked. At fault time, an
2883          * attempt will be made for VM_NORESERVE to allocate a page
2884          * and filesystem quota without using reserves
2885          */
2886         if (acctflag & VM_NORESERVE)
2887                 return 0;
2888
2889         /*
2890          * Shared mappings base their reservation on the number of pages that
2891          * are already allocated on behalf of the file. Private mappings need
2892          * to reserve the full area even if read-only as mprotect() may be
2893          * called to make the mapping read-write. Assume !vma is a shm mapping
2894          */
2895         if (!vma || vma->vm_flags & VM_MAYSHARE)
2896                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2897         else {
2898                 struct resv_map *resv_map = resv_map_alloc();
2899                 if (!resv_map)
2900                         return -ENOMEM;
2901
2902                 chg = to - from;
2903
2904                 set_vma_resv_map(vma, resv_map);
2905                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2906         }
2907
2908         if (chg < 0)
2909                 return chg;
2910
2911         /* There must be enough filesystem quota for the mapping */
2912         if (hugetlb_get_quota(inode->i_mapping, chg))
2913                 return -ENOSPC;
2914
2915         /*
2916          * Check enough hugepages are available for the reservation.
2917          * Hand back the quota if there are not
2918          */
2919         ret = hugetlb_acct_memory(h, chg);
2920         if (ret < 0) {
2921                 hugetlb_put_quota(inode->i_mapping, chg);
2922                 return ret;
2923         }
2924
2925         /*
2926          * Account for the reservations made. Shared mappings record regions
2927          * that have reservations as they are shared by multiple VMAs.
2928          * When the last VMA disappears, the region map says how much
2929          * the reservation was and the page cache tells how much of
2930          * the reservation was consumed. Private mappings are per-VMA and
2931          * only the consumed reservations are tracked. When the VMA
2932          * disappears, the original reservation is the VMA size and the
2933          * consumed reservations are stored in the map. Hence, nothing
2934          * else has to be done for private mappings here
2935          */
2936         if (!vma || vma->vm_flags & VM_MAYSHARE)
2937                 region_add(&inode->i_mapping->private_list, from, to);
2938         return 0;
2939 }
2940
2941 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2942 {
2943         struct hstate *h = hstate_inode(inode);
2944         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2945
2946         spin_lock(&inode->i_lock);
2947         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2948         spin_unlock(&inode->i_lock);
2949
2950         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2951         hugetlb_acct_memory(h, -(chg - freed));
2952 }
2953
2954 #ifdef CONFIG_MEMORY_FAILURE
2955
2956 /* Should be called in hugetlb_lock */
2957 static int is_hugepage_on_freelist(struct page *hpage)
2958 {
2959         struct page *page;
2960         struct page *tmp;
2961         struct hstate *h = page_hstate(hpage);
2962         int nid = page_to_nid(hpage);
2963
2964         list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
2965                 if (page == hpage)
2966                         return 1;
2967         return 0;
2968 }
2969
2970 /*
2971  * This function is called from memory failure code.
2972  * Assume the caller holds page lock of the head page.
2973  */
2974 int dequeue_hwpoisoned_huge_page(struct page *hpage)
2975 {
2976         struct hstate *h = page_hstate(hpage);
2977         int nid = page_to_nid(hpage);
2978         int ret = -EBUSY;
2979
2980         spin_lock(&hugetlb_lock);
2981         if (is_hugepage_on_freelist(hpage)) {
2982                 list_del(&hpage->lru);
2983                 set_page_refcounted(hpage);
2984                 h->free_huge_pages--;
2985                 h->free_huge_pages_node[nid]--;
2986                 ret = 0;
2987         }
2988         spin_unlock(&hugetlb_lock);
2989         return ret;
2990 }
2991 #endif