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