hugetlb: modular state for hugetlb page size
[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/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/cpuset.h>
16 #include <linux/mutex.h>
17
18 #include <asm/page.h>
19 #include <asm/pgtable.h>
20
21 #include <linux/hugetlb.h>
22 #include "internal.h"
23
24 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
25 unsigned long max_huge_pages;
26 unsigned long sysctl_overcommit_huge_pages;
27 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
28 unsigned long hugepages_treat_as_movable;
29
30 struct hstate default_hstate;
31
32 /*
33  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
34  */
35 static DEFINE_SPINLOCK(hugetlb_lock);
36
37 /*
38  * Region tracking -- allows tracking of reservations and instantiated pages
39  *                    across the pages in a mapping.
40  *
41  * The region data structures are protected by a combination of the mmap_sem
42  * and the hugetlb_instantion_mutex.  To access or modify a region the caller
43  * must either hold the mmap_sem for write, or the mmap_sem for read and
44  * the hugetlb_instantiation mutex:
45  *
46  *      down_write(&mm->mmap_sem);
47  * or
48  *      down_read(&mm->mmap_sem);
49  *      mutex_lock(&hugetlb_instantiation_mutex);
50  */
51 struct file_region {
52         struct list_head link;
53         long from;
54         long to;
55 };
56
57 static long region_add(struct list_head *head, long f, long t)
58 {
59         struct file_region *rg, *nrg, *trg;
60
61         /* Locate the region we are either in or before. */
62         list_for_each_entry(rg, head, link)
63                 if (f <= rg->to)
64                         break;
65
66         /* Round our left edge to the current segment if it encloses us. */
67         if (f > rg->from)
68                 f = rg->from;
69
70         /* Check for and consume any regions we now overlap with. */
71         nrg = rg;
72         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
73                 if (&rg->link == head)
74                         break;
75                 if (rg->from > t)
76                         break;
77
78                 /* If this area reaches higher then extend our area to
79                  * include it completely.  If this is not the first area
80                  * which we intend to reuse, free it. */
81                 if (rg->to > t)
82                         t = rg->to;
83                 if (rg != nrg) {
84                         list_del(&rg->link);
85                         kfree(rg);
86                 }
87         }
88         nrg->from = f;
89         nrg->to = t;
90         return 0;
91 }
92
93 static long region_chg(struct list_head *head, long f, long t)
94 {
95         struct file_region *rg, *nrg;
96         long chg = 0;
97
98         /* Locate the region we are before or in. */
99         list_for_each_entry(rg, head, link)
100                 if (f <= rg->to)
101                         break;
102
103         /* If we are below the current region then a new region is required.
104          * Subtle, allocate a new region at the position but make it zero
105          * size such that we can guarantee to record the reservation. */
106         if (&rg->link == head || t < rg->from) {
107                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
108                 if (!nrg)
109                         return -ENOMEM;
110                 nrg->from = f;
111                 nrg->to   = f;
112                 INIT_LIST_HEAD(&nrg->link);
113                 list_add(&nrg->link, rg->link.prev);
114
115                 return t - f;
116         }
117
118         /* Round our left edge to the current segment if it encloses us. */
119         if (f > rg->from)
120                 f = rg->from;
121         chg = t - f;
122
123         /* Check for and consume any regions we now overlap with. */
124         list_for_each_entry(rg, rg->link.prev, link) {
125                 if (&rg->link == head)
126                         break;
127                 if (rg->from > t)
128                         return chg;
129
130                 /* We overlap with this area, if it extends futher than
131                  * us then we must extend ourselves.  Account for its
132                  * existing reservation. */
133                 if (rg->to > t) {
134                         chg += rg->to - t;
135                         t = rg->to;
136                 }
137                 chg -= rg->to - rg->from;
138         }
139         return chg;
140 }
141
142 static long region_truncate(struct list_head *head, long end)
143 {
144         struct file_region *rg, *trg;
145         long chg = 0;
146
147         /* Locate the region we are either in or before. */
148         list_for_each_entry(rg, head, link)
149                 if (end <= rg->to)
150                         break;
151         if (&rg->link == head)
152                 return 0;
153
154         /* If we are in the middle of a region then adjust it. */
155         if (end > rg->from) {
156                 chg = rg->to - end;
157                 rg->to = end;
158                 rg = list_entry(rg->link.next, typeof(*rg), link);
159         }
160
161         /* Drop any remaining regions. */
162         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
163                 if (&rg->link == head)
164                         break;
165                 chg += rg->to - rg->from;
166                 list_del(&rg->link);
167                 kfree(rg);
168         }
169         return chg;
170 }
171
172 static long region_count(struct list_head *head, long f, long t)
173 {
174         struct file_region *rg;
175         long chg = 0;
176
177         /* Locate each segment we overlap with, and count that overlap. */
178         list_for_each_entry(rg, head, link) {
179                 int seg_from;
180                 int seg_to;
181
182                 if (rg->to <= f)
183                         continue;
184                 if (rg->from >= t)
185                         break;
186
187                 seg_from = max(rg->from, f);
188                 seg_to = min(rg->to, t);
189
190                 chg += seg_to - seg_from;
191         }
192
193         return chg;
194 }
195
196 /*
197  * Convert the address within this vma to the page offset within
198  * the mapping, in pagecache page units; huge pages here.
199  */
200 static pgoff_t vma_hugecache_offset(struct hstate *h,
201                         struct vm_area_struct *vma, unsigned long address)
202 {
203         return ((address - vma->vm_start) >> huge_page_shift(h)) +
204                         (vma->vm_pgoff >> huge_page_order(h));
205 }
206
207 /*
208  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
209  * bits of the reservation map pointer, which are always clear due to
210  * alignment.
211  */
212 #define HPAGE_RESV_OWNER    (1UL << 0)
213 #define HPAGE_RESV_UNMAPPED (1UL << 1)
214 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
215
216 /*
217  * These helpers are used to track how many pages are reserved for
218  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
219  * is guaranteed to have their future faults succeed.
220  *
221  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
222  * the reserve counters are updated with the hugetlb_lock held. It is safe
223  * to reset the VMA at fork() time as it is not in use yet and there is no
224  * chance of the global counters getting corrupted as a result of the values.
225  *
226  * The private mapping reservation is represented in a subtly different
227  * manner to a shared mapping.  A shared mapping has a region map associated
228  * with the underlying file, this region map represents the backing file
229  * pages which have ever had a reservation assigned which this persists even
230  * after the page is instantiated.  A private mapping has a region map
231  * associated with the original mmap which is attached to all VMAs which
232  * reference it, this region map represents those offsets which have consumed
233  * reservation ie. where pages have been instantiated.
234  */
235 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
236 {
237         return (unsigned long)vma->vm_private_data;
238 }
239
240 static void set_vma_private_data(struct vm_area_struct *vma,
241                                                         unsigned long value)
242 {
243         vma->vm_private_data = (void *)value;
244 }
245
246 struct resv_map {
247         struct kref refs;
248         struct list_head regions;
249 };
250
251 struct resv_map *resv_map_alloc(void)
252 {
253         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
254         if (!resv_map)
255                 return NULL;
256
257         kref_init(&resv_map->refs);
258         INIT_LIST_HEAD(&resv_map->regions);
259
260         return resv_map;
261 }
262
263 void resv_map_release(struct kref *ref)
264 {
265         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
266
267         /* Clear out any active regions before we release the map. */
268         region_truncate(&resv_map->regions, 0);
269         kfree(resv_map);
270 }
271
272 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
273 {
274         VM_BUG_ON(!is_vm_hugetlb_page(vma));
275         if (!(vma->vm_flags & VM_SHARED))
276                 return (struct resv_map *)(get_vma_private_data(vma) &
277                                                         ~HPAGE_RESV_MASK);
278         return 0;
279 }
280
281 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
282 {
283         VM_BUG_ON(!is_vm_hugetlb_page(vma));
284         VM_BUG_ON(vma->vm_flags & VM_SHARED);
285
286         set_vma_private_data(vma, (get_vma_private_data(vma) &
287                                 HPAGE_RESV_MASK) | (unsigned long)map);
288 }
289
290 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
291 {
292         VM_BUG_ON(!is_vm_hugetlb_page(vma));
293         VM_BUG_ON(vma->vm_flags & VM_SHARED);
294
295         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
296 }
297
298 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
299 {
300         VM_BUG_ON(!is_vm_hugetlb_page(vma));
301
302         return (get_vma_private_data(vma) & flag) != 0;
303 }
304
305 /* Decrement the reserved pages in the hugepage pool by one */
306 static void decrement_hugepage_resv_vma(struct hstate *h,
307                         struct vm_area_struct *vma)
308 {
309         if (vma->vm_flags & VM_NORESERVE)
310                 return;
311
312         if (vma->vm_flags & VM_SHARED) {
313                 /* Shared mappings always use reserves */
314                 h->resv_huge_pages--;
315         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
316                 /*
317                  * Only the process that called mmap() has reserves for
318                  * private mappings.
319                  */
320                 h->resv_huge_pages--;
321         }
322 }
323
324 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
325 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
326 {
327         VM_BUG_ON(!is_vm_hugetlb_page(vma));
328         if (!(vma->vm_flags & VM_SHARED))
329                 vma->vm_private_data = (void *)0;
330 }
331
332 /* Returns true if the VMA has associated reserve pages */
333 static int vma_has_private_reserves(struct vm_area_struct *vma)
334 {
335         if (vma->vm_flags & VM_SHARED)
336                 return 0;
337         if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER))
338                 return 0;
339         return 1;
340 }
341
342 static void clear_huge_page(struct page *page,
343                         unsigned long addr, unsigned long sz)
344 {
345         int i;
346
347         might_sleep();
348         for (i = 0; i < sz/PAGE_SIZE; i++) {
349                 cond_resched();
350                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
351         }
352 }
353
354 static void copy_huge_page(struct page *dst, struct page *src,
355                            unsigned long addr, struct vm_area_struct *vma)
356 {
357         int i;
358         struct hstate *h = hstate_vma(vma);
359
360         might_sleep();
361         for (i = 0; i < pages_per_huge_page(h); i++) {
362                 cond_resched();
363                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
364         }
365 }
366
367 static void enqueue_huge_page(struct hstate *h, struct page *page)
368 {
369         int nid = page_to_nid(page);
370         list_add(&page->lru, &h->hugepage_freelists[nid]);
371         h->free_huge_pages++;
372         h->free_huge_pages_node[nid]++;
373 }
374
375 static struct page *dequeue_huge_page(struct hstate *h)
376 {
377         int nid;
378         struct page *page = NULL;
379
380         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
381                 if (!list_empty(&h->hugepage_freelists[nid])) {
382                         page = list_entry(h->hugepage_freelists[nid].next,
383                                           struct page, lru);
384                         list_del(&page->lru);
385                         h->free_huge_pages--;
386                         h->free_huge_pages_node[nid]--;
387                         break;
388                 }
389         }
390         return page;
391 }
392
393 static struct page *dequeue_huge_page_vma(struct hstate *h,
394                                 struct vm_area_struct *vma,
395                                 unsigned long address, int avoid_reserve)
396 {
397         int nid;
398         struct page *page = NULL;
399         struct mempolicy *mpol;
400         nodemask_t *nodemask;
401         struct zonelist *zonelist = huge_zonelist(vma, address,
402                                         htlb_alloc_mask, &mpol, &nodemask);
403         struct zone *zone;
404         struct zoneref *z;
405
406         /*
407          * A child process with MAP_PRIVATE mappings created by their parent
408          * have no page reserves. This check ensures that reservations are
409          * not "stolen". The child may still get SIGKILLed
410          */
411         if (!vma_has_private_reserves(vma) &&
412                         h->free_huge_pages - h->resv_huge_pages == 0)
413                 return NULL;
414
415         /* If reserves cannot be used, ensure enough pages are in the pool */
416         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
417                 return NULL;
418
419         for_each_zone_zonelist_nodemask(zone, z, zonelist,
420                                                 MAX_NR_ZONES - 1, nodemask) {
421                 nid = zone_to_nid(zone);
422                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
423                     !list_empty(&h->hugepage_freelists[nid])) {
424                         page = list_entry(h->hugepage_freelists[nid].next,
425                                           struct page, lru);
426                         list_del(&page->lru);
427                         h->free_huge_pages--;
428                         h->free_huge_pages_node[nid]--;
429
430                         if (!avoid_reserve)
431                                 decrement_hugepage_resv_vma(h, vma);
432
433                         break;
434                 }
435         }
436         mpol_cond_put(mpol);
437         return page;
438 }
439
440 static void update_and_free_page(struct hstate *h, struct page *page)
441 {
442         int i;
443
444         h->nr_huge_pages--;
445         h->nr_huge_pages_node[page_to_nid(page)]--;
446         for (i = 0; i < pages_per_huge_page(h); i++) {
447                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
448                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
449                                 1 << PG_private | 1<< PG_writeback);
450         }
451         set_compound_page_dtor(page, NULL);
452         set_page_refcounted(page);
453         arch_release_hugepage(page);
454         __free_pages(page, huge_page_order(h));
455 }
456
457 static void free_huge_page(struct page *page)
458 {
459         /*
460          * Can't pass hstate in here because it is called from the
461          * compound page destructor.
462          */
463         struct hstate *h = &default_hstate;
464         int nid = page_to_nid(page);
465         struct address_space *mapping;
466
467         mapping = (struct address_space *) page_private(page);
468         set_page_private(page, 0);
469         BUG_ON(page_count(page));
470         INIT_LIST_HEAD(&page->lru);
471
472         spin_lock(&hugetlb_lock);
473         if (h->surplus_huge_pages_node[nid]) {
474                 update_and_free_page(h, page);
475                 h->surplus_huge_pages--;
476                 h->surplus_huge_pages_node[nid]--;
477         } else {
478                 enqueue_huge_page(h, page);
479         }
480         spin_unlock(&hugetlb_lock);
481         if (mapping)
482                 hugetlb_put_quota(mapping, 1);
483 }
484
485 /*
486  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
487  * balanced by operating on them in a round-robin fashion.
488  * Returns 1 if an adjustment was made.
489  */
490 static int adjust_pool_surplus(struct hstate *h, int delta)
491 {
492         static int prev_nid;
493         int nid = prev_nid;
494         int ret = 0;
495
496         VM_BUG_ON(delta != -1 && delta != 1);
497         do {
498                 nid = next_node(nid, node_online_map);
499                 if (nid == MAX_NUMNODES)
500                         nid = first_node(node_online_map);
501
502                 /* To shrink on this node, there must be a surplus page */
503                 if (delta < 0 && !h->surplus_huge_pages_node[nid])
504                         continue;
505                 /* Surplus cannot exceed the total number of pages */
506                 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
507                                                 h->nr_huge_pages_node[nid])
508                         continue;
509
510                 h->surplus_huge_pages += delta;
511                 h->surplus_huge_pages_node[nid] += delta;
512                 ret = 1;
513                 break;
514         } while (nid != prev_nid);
515
516         prev_nid = nid;
517         return ret;
518 }
519
520 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
521 {
522         set_compound_page_dtor(page, free_huge_page);
523         spin_lock(&hugetlb_lock);
524         h->nr_huge_pages++;
525         h->nr_huge_pages_node[nid]++;
526         spin_unlock(&hugetlb_lock);
527         put_page(page); /* free it into the hugepage allocator */
528 }
529
530 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
531 {
532         struct page *page;
533
534         page = alloc_pages_node(nid,
535                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
536                                                 __GFP_REPEAT|__GFP_NOWARN,
537                 huge_page_order(h));
538         if (page) {
539                 if (arch_prepare_hugepage(page)) {
540                         __free_pages(page, HUGETLB_PAGE_ORDER);
541                         return NULL;
542                 }
543                 prep_new_huge_page(h, page, nid);
544         }
545
546         return page;
547 }
548
549 static int alloc_fresh_huge_page(struct hstate *h)
550 {
551         struct page *page;
552         int start_nid;
553         int next_nid;
554         int ret = 0;
555
556         start_nid = h->hugetlb_next_nid;
557
558         do {
559                 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
560                 if (page)
561                         ret = 1;
562                 /*
563                  * Use a helper variable to find the next node and then
564                  * copy it back to hugetlb_next_nid afterwards:
565                  * otherwise there's a window in which a racer might
566                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
567                  * But we don't need to use a spin_lock here: it really
568                  * doesn't matter if occasionally a racer chooses the
569                  * same nid as we do.  Move nid forward in the mask even
570                  * if we just successfully allocated a hugepage so that
571                  * the next caller gets hugepages on the next node.
572                  */
573                 next_nid = next_node(h->hugetlb_next_nid, node_online_map);
574                 if (next_nid == MAX_NUMNODES)
575                         next_nid = first_node(node_online_map);
576                 h->hugetlb_next_nid = next_nid;
577         } while (!page && h->hugetlb_next_nid != start_nid);
578
579         if (ret)
580                 count_vm_event(HTLB_BUDDY_PGALLOC);
581         else
582                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
583
584         return ret;
585 }
586
587 static struct page *alloc_buddy_huge_page(struct hstate *h,
588                         struct vm_area_struct *vma, unsigned long address)
589 {
590         struct page *page;
591         unsigned int nid;
592
593         /*
594          * Assume we will successfully allocate the surplus page to
595          * prevent racing processes from causing the surplus to exceed
596          * overcommit
597          *
598          * This however introduces a different race, where a process B
599          * tries to grow the static hugepage pool while alloc_pages() is
600          * called by process A. B will only examine the per-node
601          * counters in determining if surplus huge pages can be
602          * converted to normal huge pages in adjust_pool_surplus(). A
603          * won't be able to increment the per-node counter, until the
604          * lock is dropped by B, but B doesn't drop hugetlb_lock until
605          * no more huge pages can be converted from surplus to normal
606          * state (and doesn't try to convert again). Thus, we have a
607          * case where a surplus huge page exists, the pool is grown, and
608          * the surplus huge page still exists after, even though it
609          * should just have been converted to a normal huge page. This
610          * does not leak memory, though, as the hugepage will be freed
611          * once it is out of use. It also does not allow the counters to
612          * go out of whack in adjust_pool_surplus() as we don't modify
613          * the node values until we've gotten the hugepage and only the
614          * per-node value is checked there.
615          */
616         spin_lock(&hugetlb_lock);
617         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
618                 spin_unlock(&hugetlb_lock);
619                 return NULL;
620         } else {
621                 h->nr_huge_pages++;
622                 h->surplus_huge_pages++;
623         }
624         spin_unlock(&hugetlb_lock);
625
626         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
627                                         __GFP_REPEAT|__GFP_NOWARN,
628                                         huge_page_order(h));
629
630         spin_lock(&hugetlb_lock);
631         if (page) {
632                 /*
633                  * This page is now managed by the hugetlb allocator and has
634                  * no users -- drop the buddy allocator's reference.
635                  */
636                 put_page_testzero(page);
637                 VM_BUG_ON(page_count(page));
638                 nid = page_to_nid(page);
639                 set_compound_page_dtor(page, free_huge_page);
640                 /*
641                  * We incremented the global counters already
642                  */
643                 h->nr_huge_pages_node[nid]++;
644                 h->surplus_huge_pages_node[nid]++;
645                 __count_vm_event(HTLB_BUDDY_PGALLOC);
646         } else {
647                 h->nr_huge_pages--;
648                 h->surplus_huge_pages--;
649                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
650         }
651         spin_unlock(&hugetlb_lock);
652
653         return page;
654 }
655
656 /*
657  * Increase the hugetlb pool such that it can accomodate a reservation
658  * of size 'delta'.
659  */
660 static int gather_surplus_pages(struct hstate *h, int delta)
661 {
662         struct list_head surplus_list;
663         struct page *page, *tmp;
664         int ret, i;
665         int needed, allocated;
666
667         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
668         if (needed <= 0) {
669                 h->resv_huge_pages += delta;
670                 return 0;
671         }
672
673         allocated = 0;
674         INIT_LIST_HEAD(&surplus_list);
675
676         ret = -ENOMEM;
677 retry:
678         spin_unlock(&hugetlb_lock);
679         for (i = 0; i < needed; i++) {
680                 page = alloc_buddy_huge_page(h, NULL, 0);
681                 if (!page) {
682                         /*
683                          * We were not able to allocate enough pages to
684                          * satisfy the entire reservation so we free what
685                          * we've allocated so far.
686                          */
687                         spin_lock(&hugetlb_lock);
688                         needed = 0;
689                         goto free;
690                 }
691
692                 list_add(&page->lru, &surplus_list);
693         }
694         allocated += needed;
695
696         /*
697          * After retaking hugetlb_lock, we need to recalculate 'needed'
698          * because either resv_huge_pages or free_huge_pages may have changed.
699          */
700         spin_lock(&hugetlb_lock);
701         needed = (h->resv_huge_pages + delta) -
702                         (h->free_huge_pages + allocated);
703         if (needed > 0)
704                 goto retry;
705
706         /*
707          * The surplus_list now contains _at_least_ the number of extra pages
708          * needed to accomodate the reservation.  Add the appropriate number
709          * of pages to the hugetlb pool and free the extras back to the buddy
710          * allocator.  Commit the entire reservation here to prevent another
711          * process from stealing the pages as they are added to the pool but
712          * before they are reserved.
713          */
714         needed += allocated;
715         h->resv_huge_pages += delta;
716         ret = 0;
717 free:
718         /* Free the needed pages to the hugetlb pool */
719         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
720                 if ((--needed) < 0)
721                         break;
722                 list_del(&page->lru);
723                 enqueue_huge_page(h, page);
724         }
725
726         /* Free unnecessary surplus pages to the buddy allocator */
727         if (!list_empty(&surplus_list)) {
728                 spin_unlock(&hugetlb_lock);
729                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
730                         list_del(&page->lru);
731                         /*
732                          * The page has a reference count of zero already, so
733                          * call free_huge_page directly instead of using
734                          * put_page.  This must be done with hugetlb_lock
735                          * unlocked which is safe because free_huge_page takes
736                          * hugetlb_lock before deciding how to free the page.
737                          */
738                         free_huge_page(page);
739                 }
740                 spin_lock(&hugetlb_lock);
741         }
742
743         return ret;
744 }
745
746 /*
747  * When releasing a hugetlb pool reservation, any surplus pages that were
748  * allocated to satisfy the reservation must be explicitly freed if they were
749  * never used.
750  */
751 static void return_unused_surplus_pages(struct hstate *h,
752                                         unsigned long unused_resv_pages)
753 {
754         static int nid = -1;
755         struct page *page;
756         unsigned long nr_pages;
757
758         /*
759          * We want to release as many surplus pages as possible, spread
760          * evenly across all nodes. Iterate across all nodes until we
761          * can no longer free unreserved surplus pages. This occurs when
762          * the nodes with surplus pages have no free pages.
763          */
764         unsigned long remaining_iterations = num_online_nodes();
765
766         /* Uncommit the reservation */
767         h->resv_huge_pages -= unused_resv_pages;
768
769         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
770
771         while (remaining_iterations-- && nr_pages) {
772                 nid = next_node(nid, node_online_map);
773                 if (nid == MAX_NUMNODES)
774                         nid = first_node(node_online_map);
775
776                 if (!h->surplus_huge_pages_node[nid])
777                         continue;
778
779                 if (!list_empty(&h->hugepage_freelists[nid])) {
780                         page = list_entry(h->hugepage_freelists[nid].next,
781                                           struct page, lru);
782                         list_del(&page->lru);
783                         update_and_free_page(h, page);
784                         h->free_huge_pages--;
785                         h->free_huge_pages_node[nid]--;
786                         h->surplus_huge_pages--;
787                         h->surplus_huge_pages_node[nid]--;
788                         nr_pages--;
789                         remaining_iterations = num_online_nodes();
790                 }
791         }
792 }
793
794 /*
795  * Determine if the huge page at addr within the vma has an associated
796  * reservation.  Where it does not we will need to logically increase
797  * reservation and actually increase quota before an allocation can occur.
798  * Where any new reservation would be required the reservation change is
799  * prepared, but not committed.  Once the page has been quota'd allocated
800  * an instantiated the change should be committed via vma_commit_reservation.
801  * No action is required on failure.
802  */
803 static int vma_needs_reservation(struct hstate *h,
804                         struct vm_area_struct *vma, unsigned long addr)
805 {
806         struct address_space *mapping = vma->vm_file->f_mapping;
807         struct inode *inode = mapping->host;
808
809         if (vma->vm_flags & VM_SHARED) {
810                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
811                 return region_chg(&inode->i_mapping->private_list,
812                                                         idx, idx + 1);
813
814         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
815                 return 1;
816
817         } else  {
818                 int err;
819                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
820                 struct resv_map *reservations = vma_resv_map(vma);
821
822                 err = region_chg(&reservations->regions, idx, idx + 1);
823                 if (err < 0)
824                         return err;
825                 return 0;
826         }
827 }
828 static void vma_commit_reservation(struct hstate *h,
829                         struct vm_area_struct *vma, unsigned long addr)
830 {
831         struct address_space *mapping = vma->vm_file->f_mapping;
832         struct inode *inode = mapping->host;
833
834         if (vma->vm_flags & VM_SHARED) {
835                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
836                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
837
838         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
839                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
840                 struct resv_map *reservations = vma_resv_map(vma);
841
842                 /* Mark this page used in the map. */
843                 region_add(&reservations->regions, idx, idx + 1);
844         }
845 }
846
847 static struct page *alloc_huge_page(struct vm_area_struct *vma,
848                                     unsigned long addr, int avoid_reserve)
849 {
850         struct hstate *h = hstate_vma(vma);
851         struct page *page;
852         struct address_space *mapping = vma->vm_file->f_mapping;
853         struct inode *inode = mapping->host;
854         unsigned int chg;
855
856         /*
857          * Processes that did not create the mapping will have no reserves and
858          * will not have accounted against quota. Check that the quota can be
859          * made before satisfying the allocation
860          * MAP_NORESERVE mappings may also need pages and quota allocated
861          * if no reserve mapping overlaps.
862          */
863         chg = vma_needs_reservation(h, vma, addr);
864         if (chg < 0)
865                 return ERR_PTR(chg);
866         if (chg)
867                 if (hugetlb_get_quota(inode->i_mapping, chg))
868                         return ERR_PTR(-ENOSPC);
869
870         spin_lock(&hugetlb_lock);
871         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
872         spin_unlock(&hugetlb_lock);
873
874         if (!page) {
875                 page = alloc_buddy_huge_page(h, vma, addr);
876                 if (!page) {
877                         hugetlb_put_quota(inode->i_mapping, chg);
878                         return ERR_PTR(-VM_FAULT_OOM);
879                 }
880         }
881
882         set_page_refcounted(page);
883         set_page_private(page, (unsigned long) mapping);
884
885         vma_commit_reservation(h, vma, addr);
886
887         return page;
888 }
889
890 static int __init hugetlb_init(void)
891 {
892         unsigned long i;
893         struct hstate *h = &default_hstate;
894
895         if (HPAGE_SHIFT == 0)
896                 return 0;
897
898         if (!h->order) {
899                 h->order = HPAGE_SHIFT - PAGE_SHIFT;
900                 h->mask = HPAGE_MASK;
901         }
902
903         for (i = 0; i < MAX_NUMNODES; ++i)
904                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
905
906         h->hugetlb_next_nid = first_node(node_online_map);
907
908         for (i = 0; i < max_huge_pages; ++i) {
909                 if (!alloc_fresh_huge_page(h))
910                         break;
911         }
912         max_huge_pages = h->free_huge_pages = h->nr_huge_pages = i;
913         printk(KERN_INFO "Total HugeTLB memory allocated, %ld\n",
914                         h->free_huge_pages);
915         return 0;
916 }
917 module_init(hugetlb_init);
918
919 static int __init hugetlb_setup(char *s)
920 {
921         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
922                 max_huge_pages = 0;
923         return 1;
924 }
925 __setup("hugepages=", hugetlb_setup);
926
927 static unsigned int cpuset_mems_nr(unsigned int *array)
928 {
929         int node;
930         unsigned int nr = 0;
931
932         for_each_node_mask(node, cpuset_current_mems_allowed)
933                 nr += array[node];
934
935         return nr;
936 }
937
938 #ifdef CONFIG_SYSCTL
939 #ifdef CONFIG_HIGHMEM
940 static void try_to_free_low(struct hstate *h, unsigned long count)
941 {
942         int i;
943
944         for (i = 0; i < MAX_NUMNODES; ++i) {
945                 struct page *page, *next;
946                 struct list_head *freel = &h->hugepage_freelists[i];
947                 list_for_each_entry_safe(page, next, freel, lru) {
948                         if (count >= h->nr_huge_pages)
949                                 return;
950                         if (PageHighMem(page))
951                                 continue;
952                         list_del(&page->lru);
953                         update_and_free_page(page);
954                         h->free_huge_pages--;
955                         h->free_huge_pages_node[page_to_nid(page)]--;
956                 }
957         }
958 }
959 #else
960 static inline void try_to_free_low(struct hstate *h, unsigned long count)
961 {
962 }
963 #endif
964
965 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
966 static unsigned long set_max_huge_pages(unsigned long count)
967 {
968         unsigned long min_count, ret;
969         struct hstate *h = &default_hstate;
970
971         /*
972          * Increase the pool size
973          * First take pages out of surplus state.  Then make up the
974          * remaining difference by allocating fresh huge pages.
975          *
976          * We might race with alloc_buddy_huge_page() here and be unable
977          * to convert a surplus huge page to a normal huge page. That is
978          * not critical, though, it just means the overall size of the
979          * pool might be one hugepage larger than it needs to be, but
980          * within all the constraints specified by the sysctls.
981          */
982         spin_lock(&hugetlb_lock);
983         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
984                 if (!adjust_pool_surplus(h, -1))
985                         break;
986         }
987
988         while (count > persistent_huge_pages(h)) {
989                 /*
990                  * If this allocation races such that we no longer need the
991                  * page, free_huge_page will handle it by freeing the page
992                  * and reducing the surplus.
993                  */
994                 spin_unlock(&hugetlb_lock);
995                 ret = alloc_fresh_huge_page(h);
996                 spin_lock(&hugetlb_lock);
997                 if (!ret)
998                         goto out;
999
1000         }
1001
1002         /*
1003          * Decrease the pool size
1004          * First return free pages to the buddy allocator (being careful
1005          * to keep enough around to satisfy reservations).  Then place
1006          * pages into surplus state as needed so the pool will shrink
1007          * to the desired size as pages become free.
1008          *
1009          * By placing pages into the surplus state independent of the
1010          * overcommit value, we are allowing the surplus pool size to
1011          * exceed overcommit. There are few sane options here. Since
1012          * alloc_buddy_huge_page() is checking the global counter,
1013          * though, we'll note that we're not allowed to exceed surplus
1014          * and won't grow the pool anywhere else. Not until one of the
1015          * sysctls are changed, or the surplus pages go out of use.
1016          */
1017         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1018         min_count = max(count, min_count);
1019         try_to_free_low(h, min_count);
1020         while (min_count < persistent_huge_pages(h)) {
1021                 struct page *page = dequeue_huge_page(h);
1022                 if (!page)
1023                         break;
1024                 update_and_free_page(h, page);
1025         }
1026         while (count < persistent_huge_pages(h)) {
1027                 if (!adjust_pool_surplus(h, 1))
1028                         break;
1029         }
1030 out:
1031         ret = persistent_huge_pages(h);
1032         spin_unlock(&hugetlb_lock);
1033         return ret;
1034 }
1035
1036 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1037                            struct file *file, void __user *buffer,
1038                            size_t *length, loff_t *ppos)
1039 {
1040         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1041         max_huge_pages = set_max_huge_pages(max_huge_pages);
1042         return 0;
1043 }
1044
1045 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1046                         struct file *file, void __user *buffer,
1047                         size_t *length, loff_t *ppos)
1048 {
1049         proc_dointvec(table, write, file, buffer, length, ppos);
1050         if (hugepages_treat_as_movable)
1051                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1052         else
1053                 htlb_alloc_mask = GFP_HIGHUSER;
1054         return 0;
1055 }
1056
1057 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1058                         struct file *file, void __user *buffer,
1059                         size_t *length, loff_t *ppos)
1060 {
1061         struct hstate *h = &default_hstate;
1062         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1063         spin_lock(&hugetlb_lock);
1064         h->nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
1065         spin_unlock(&hugetlb_lock);
1066         return 0;
1067 }
1068
1069 #endif /* CONFIG_SYSCTL */
1070
1071 int hugetlb_report_meminfo(char *buf)
1072 {
1073         struct hstate *h = &default_hstate;
1074         return sprintf(buf,
1075                         "HugePages_Total: %5lu\n"
1076                         "HugePages_Free:  %5lu\n"
1077                         "HugePages_Rsvd:  %5lu\n"
1078                         "HugePages_Surp:  %5lu\n"
1079                         "Hugepagesize:    %5lu kB\n",
1080                         h->nr_huge_pages,
1081                         h->free_huge_pages,
1082                         h->resv_huge_pages,
1083                         h->surplus_huge_pages,
1084                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1085 }
1086
1087 int hugetlb_report_node_meminfo(int nid, char *buf)
1088 {
1089         struct hstate *h = &default_hstate;
1090         return sprintf(buf,
1091                 "Node %d HugePages_Total: %5u\n"
1092                 "Node %d HugePages_Free:  %5u\n"
1093                 "Node %d HugePages_Surp:  %5u\n",
1094                 nid, h->nr_huge_pages_node[nid],
1095                 nid, h->free_huge_pages_node[nid],
1096                 nid, h->surplus_huge_pages_node[nid]);
1097 }
1098
1099 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1100 unsigned long hugetlb_total_pages(void)
1101 {
1102         struct hstate *h = &default_hstate;
1103         return h->nr_huge_pages * pages_per_huge_page(h);
1104 }
1105
1106 static int hugetlb_acct_memory(struct hstate *h, long delta)
1107 {
1108         int ret = -ENOMEM;
1109
1110         spin_lock(&hugetlb_lock);
1111         /*
1112          * When cpuset is configured, it breaks the strict hugetlb page
1113          * reservation as the accounting is done on a global variable. Such
1114          * reservation is completely rubbish in the presence of cpuset because
1115          * the reservation is not checked against page availability for the
1116          * current cpuset. Application can still potentially OOM'ed by kernel
1117          * with lack of free htlb page in cpuset that the task is in.
1118          * Attempt to enforce strict accounting with cpuset is almost
1119          * impossible (or too ugly) because cpuset is too fluid that
1120          * task or memory node can be dynamically moved between cpusets.
1121          *
1122          * The change of semantics for shared hugetlb mapping with cpuset is
1123          * undesirable. However, in order to preserve some of the semantics,
1124          * we fall back to check against current free page availability as
1125          * a best attempt and hopefully to minimize the impact of changing
1126          * semantics that cpuset has.
1127          */
1128         if (delta > 0) {
1129                 if (gather_surplus_pages(h, delta) < 0)
1130                         goto out;
1131
1132                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1133                         return_unused_surplus_pages(h, delta);
1134                         goto out;
1135                 }
1136         }
1137
1138         ret = 0;
1139         if (delta < 0)
1140                 return_unused_surplus_pages(h, (unsigned long) -delta);
1141
1142 out:
1143         spin_unlock(&hugetlb_lock);
1144         return ret;
1145 }
1146
1147 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1148 {
1149         struct resv_map *reservations = vma_resv_map(vma);
1150
1151         /*
1152          * This new VMA should share its siblings reservation map if present.
1153          * The VMA will only ever have a valid reservation map pointer where
1154          * it is being copied for another still existing VMA.  As that VMA
1155          * has a reference to the reservation map it cannot dissappear until
1156          * after this open call completes.  It is therefore safe to take a
1157          * new reference here without additional locking.
1158          */
1159         if (reservations)
1160                 kref_get(&reservations->refs);
1161 }
1162
1163 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1164 {
1165         struct hstate *h = hstate_vma(vma);
1166         struct resv_map *reservations = vma_resv_map(vma);
1167         unsigned long reserve;
1168         unsigned long start;
1169         unsigned long end;
1170
1171         if (reservations) {
1172                 start = vma_hugecache_offset(h, vma, vma->vm_start);
1173                 end = vma_hugecache_offset(h, vma, vma->vm_end);
1174
1175                 reserve = (end - start) -
1176                         region_count(&reservations->regions, start, end);
1177
1178                 kref_put(&reservations->refs, resv_map_release);
1179
1180                 if (reserve)
1181                         hugetlb_acct_memory(h, -reserve);
1182         }
1183 }
1184
1185 /*
1186  * We cannot handle pagefaults against hugetlb pages at all.  They cause
1187  * handle_mm_fault() to try to instantiate regular-sized pages in the
1188  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
1189  * this far.
1190  */
1191 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1192 {
1193         BUG();
1194         return 0;
1195 }
1196
1197 struct vm_operations_struct hugetlb_vm_ops = {
1198         .fault = hugetlb_vm_op_fault,
1199         .open = hugetlb_vm_op_open,
1200         .close = hugetlb_vm_op_close,
1201 };
1202
1203 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1204                                 int writable)
1205 {
1206         pte_t entry;
1207
1208         if (writable) {
1209                 entry =
1210                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1211         } else {
1212                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1213         }
1214         entry = pte_mkyoung(entry);
1215         entry = pte_mkhuge(entry);
1216
1217         return entry;
1218 }
1219
1220 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1221                                    unsigned long address, pte_t *ptep)
1222 {
1223         pte_t entry;
1224
1225         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1226         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1227                 update_mmu_cache(vma, address, entry);
1228         }
1229 }
1230
1231
1232 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1233                             struct vm_area_struct *vma)
1234 {
1235         pte_t *src_pte, *dst_pte, entry;
1236         struct page *ptepage;
1237         unsigned long addr;
1238         int cow;
1239         struct hstate *h = hstate_vma(vma);
1240         unsigned long sz = huge_page_size(h);
1241
1242         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1243
1244         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1245                 src_pte = huge_pte_offset(src, addr);
1246                 if (!src_pte)
1247                         continue;
1248                 dst_pte = huge_pte_alloc(dst, addr, sz);
1249                 if (!dst_pte)
1250                         goto nomem;
1251
1252                 /* If the pagetables are shared don't copy or take references */
1253                 if (dst_pte == src_pte)
1254                         continue;
1255
1256                 spin_lock(&dst->page_table_lock);
1257                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1258                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1259                         if (cow)
1260                                 huge_ptep_set_wrprotect(src, addr, src_pte);
1261                         entry = huge_ptep_get(src_pte);
1262                         ptepage = pte_page(entry);
1263                         get_page(ptepage);
1264                         set_huge_pte_at(dst, addr, dst_pte, entry);
1265                 }
1266                 spin_unlock(&src->page_table_lock);
1267                 spin_unlock(&dst->page_table_lock);
1268         }
1269         return 0;
1270
1271 nomem:
1272         return -ENOMEM;
1273 }
1274
1275 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1276                             unsigned long end, struct page *ref_page)
1277 {
1278         struct mm_struct *mm = vma->vm_mm;
1279         unsigned long address;
1280         pte_t *ptep;
1281         pte_t pte;
1282         struct page *page;
1283         struct page *tmp;
1284         struct hstate *h = hstate_vma(vma);
1285         unsigned long sz = huge_page_size(h);
1286
1287         /*
1288          * A page gathering list, protected by per file i_mmap_lock. The
1289          * lock is used to avoid list corruption from multiple unmapping
1290          * of the same page since we are using page->lru.
1291          */
1292         LIST_HEAD(page_list);
1293
1294         WARN_ON(!is_vm_hugetlb_page(vma));
1295         BUG_ON(start & ~huge_page_mask(h));
1296         BUG_ON(end & ~huge_page_mask(h));
1297
1298         spin_lock(&mm->page_table_lock);
1299         for (address = start; address < end; address += sz) {
1300                 ptep = huge_pte_offset(mm, address);
1301                 if (!ptep)
1302                         continue;
1303
1304                 if (huge_pmd_unshare(mm, &address, ptep))
1305                         continue;
1306
1307                 /*
1308                  * If a reference page is supplied, it is because a specific
1309                  * page is being unmapped, not a range. Ensure the page we
1310                  * are about to unmap is the actual page of interest.
1311                  */
1312                 if (ref_page) {
1313                         pte = huge_ptep_get(ptep);
1314                         if (huge_pte_none(pte))
1315                                 continue;
1316                         page = pte_page(pte);
1317                         if (page != ref_page)
1318                                 continue;
1319
1320                         /*
1321                          * Mark the VMA as having unmapped its page so that
1322                          * future faults in this VMA will fail rather than
1323                          * looking like data was lost
1324                          */
1325                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1326                 }
1327
1328                 pte = huge_ptep_get_and_clear(mm, address, ptep);
1329                 if (huge_pte_none(pte))
1330                         continue;
1331
1332                 page = pte_page(pte);
1333                 if (pte_dirty(pte))
1334                         set_page_dirty(page);
1335                 list_add(&page->lru, &page_list);
1336         }
1337         spin_unlock(&mm->page_table_lock);
1338         flush_tlb_range(vma, start, end);
1339         list_for_each_entry_safe(page, tmp, &page_list, lru) {
1340                 list_del(&page->lru);
1341                 put_page(page);
1342         }
1343 }
1344
1345 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1346                           unsigned long end, struct page *ref_page)
1347 {
1348         /*
1349          * It is undesirable to test vma->vm_file as it should be non-null
1350          * for valid hugetlb area. However, vm_file will be NULL in the error
1351          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
1352          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
1353          * to clean up. Since no pte has actually been setup, it is safe to
1354          * do nothing in this case.
1355          */
1356         if (vma->vm_file) {
1357                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1358                 __unmap_hugepage_range(vma, start, end, ref_page);
1359                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1360         }
1361 }
1362
1363 /*
1364  * This is called when the original mapper is failing to COW a MAP_PRIVATE
1365  * mappping it owns the reserve page for. The intention is to unmap the page
1366  * from other VMAs and let the children be SIGKILLed if they are faulting the
1367  * same region.
1368  */
1369 int unmap_ref_private(struct mm_struct *mm,
1370                                         struct vm_area_struct *vma,
1371                                         struct page *page,
1372                                         unsigned long address)
1373 {
1374         struct vm_area_struct *iter_vma;
1375         struct address_space *mapping;
1376         struct prio_tree_iter iter;
1377         pgoff_t pgoff;
1378
1379         /*
1380          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1381          * from page cache lookup which is in HPAGE_SIZE units.
1382          */
1383         address = address & huge_page_mask(hstate_vma(vma));
1384         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1385                 + (vma->vm_pgoff >> PAGE_SHIFT);
1386         mapping = (struct address_space *)page_private(page);
1387
1388         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1389                 /* Do not unmap the current VMA */
1390                 if (iter_vma == vma)
1391                         continue;
1392
1393                 /*
1394                  * Unmap the page from other VMAs without their own reserves.
1395                  * They get marked to be SIGKILLed if they fault in these
1396                  * areas. This is because a future no-page fault on this VMA
1397                  * could insert a zeroed page instead of the data existing
1398                  * from the time of fork. This would look like data corruption
1399                  */
1400                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1401                         unmap_hugepage_range(iter_vma,
1402                                 address, address + HPAGE_SIZE,
1403                                 page);
1404         }
1405
1406         return 1;
1407 }
1408
1409 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1410                         unsigned long address, pte_t *ptep, pte_t pte,
1411                         struct page *pagecache_page)
1412 {
1413         struct hstate *h = hstate_vma(vma);
1414         struct page *old_page, *new_page;
1415         int avoidcopy;
1416         int outside_reserve = 0;
1417
1418         old_page = pte_page(pte);
1419
1420 retry_avoidcopy:
1421         /* If no-one else is actually using this page, avoid the copy
1422          * and just make the page writable */
1423         avoidcopy = (page_count(old_page) == 1);
1424         if (avoidcopy) {
1425                 set_huge_ptep_writable(vma, address, ptep);
1426                 return 0;
1427         }
1428
1429         /*
1430          * If the process that created a MAP_PRIVATE mapping is about to
1431          * perform a COW due to a shared page count, attempt to satisfy
1432          * the allocation without using the existing reserves. The pagecache
1433          * page is used to determine if the reserve at this address was
1434          * consumed or not. If reserves were used, a partial faulted mapping
1435          * at the time of fork() could consume its reserves on COW instead
1436          * of the full address range.
1437          */
1438         if (!(vma->vm_flags & VM_SHARED) &&
1439                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1440                         old_page != pagecache_page)
1441                 outside_reserve = 1;
1442
1443         page_cache_get(old_page);
1444         new_page = alloc_huge_page(vma, address, outside_reserve);
1445
1446         if (IS_ERR(new_page)) {
1447                 page_cache_release(old_page);
1448
1449                 /*
1450                  * If a process owning a MAP_PRIVATE mapping fails to COW,
1451                  * it is due to references held by a child and an insufficient
1452                  * huge page pool. To guarantee the original mappers
1453                  * reliability, unmap the page from child processes. The child
1454                  * may get SIGKILLed if it later faults.
1455                  */
1456                 if (outside_reserve) {
1457                         BUG_ON(huge_pte_none(pte));
1458                         if (unmap_ref_private(mm, vma, old_page, address)) {
1459                                 BUG_ON(page_count(old_page) != 1);
1460                                 BUG_ON(huge_pte_none(pte));
1461                                 goto retry_avoidcopy;
1462                         }
1463                         WARN_ON_ONCE(1);
1464                 }
1465
1466                 return -PTR_ERR(new_page);
1467         }
1468
1469         spin_unlock(&mm->page_table_lock);
1470         copy_huge_page(new_page, old_page, address, vma);
1471         __SetPageUptodate(new_page);
1472         spin_lock(&mm->page_table_lock);
1473
1474         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1475         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1476                 /* Break COW */
1477                 huge_ptep_clear_flush(vma, address, ptep);
1478                 set_huge_pte_at(mm, address, ptep,
1479                                 make_huge_pte(vma, new_page, 1));
1480                 /* Make the old page be freed below */
1481                 new_page = old_page;
1482         }
1483         page_cache_release(new_page);
1484         page_cache_release(old_page);
1485         return 0;
1486 }
1487
1488 /* Return the pagecache page at a given address within a VMA */
1489 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1490                         struct vm_area_struct *vma, unsigned long address)
1491 {
1492         struct address_space *mapping;
1493         pgoff_t idx;
1494
1495         mapping = vma->vm_file->f_mapping;
1496         idx = vma_hugecache_offset(h, vma, address);
1497
1498         return find_lock_page(mapping, idx);
1499 }
1500
1501 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1502                         unsigned long address, pte_t *ptep, int write_access)
1503 {
1504         struct hstate *h = hstate_vma(vma);
1505         int ret = VM_FAULT_SIGBUS;
1506         pgoff_t idx;
1507         unsigned long size;
1508         struct page *page;
1509         struct address_space *mapping;
1510         pte_t new_pte;
1511
1512         /*
1513          * Currently, we are forced to kill the process in the event the
1514          * original mapper has unmapped pages from the child due to a failed
1515          * COW. Warn that such a situation has occured as it may not be obvious
1516          */
1517         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1518                 printk(KERN_WARNING
1519                         "PID %d killed due to inadequate hugepage pool\n",
1520                         current->pid);
1521                 return ret;
1522         }
1523
1524         mapping = vma->vm_file->f_mapping;
1525         idx = vma_hugecache_offset(h, vma, address);
1526
1527         /*
1528          * Use page lock to guard against racing truncation
1529          * before we get page_table_lock.
1530          */
1531 retry:
1532         page = find_lock_page(mapping, idx);
1533         if (!page) {
1534                 size = i_size_read(mapping->host) >> huge_page_shift(h);
1535                 if (idx >= size)
1536                         goto out;
1537                 page = alloc_huge_page(vma, address, 0);
1538                 if (IS_ERR(page)) {
1539                         ret = -PTR_ERR(page);
1540                         goto out;
1541                 }
1542                 clear_huge_page(page, address, huge_page_size(h));
1543                 __SetPageUptodate(page);
1544
1545                 if (vma->vm_flags & VM_SHARED) {
1546                         int err;
1547                         struct inode *inode = mapping->host;
1548
1549                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1550                         if (err) {
1551                                 put_page(page);
1552                                 if (err == -EEXIST)
1553                                         goto retry;
1554                                 goto out;
1555                         }
1556
1557                         spin_lock(&inode->i_lock);
1558                         inode->i_blocks += blocks_per_huge_page(h);
1559                         spin_unlock(&inode->i_lock);
1560                 } else
1561                         lock_page(page);
1562         }
1563
1564         spin_lock(&mm->page_table_lock);
1565         size = i_size_read(mapping->host) >> huge_page_shift(h);
1566         if (idx >= size)
1567                 goto backout;
1568
1569         ret = 0;
1570         if (!huge_pte_none(huge_ptep_get(ptep)))
1571                 goto backout;
1572
1573         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
1574                                 && (vma->vm_flags & VM_SHARED)));
1575         set_huge_pte_at(mm, address, ptep, new_pte);
1576
1577         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1578                 /* Optimization, do the COW without a second fault */
1579                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
1580         }
1581
1582         spin_unlock(&mm->page_table_lock);
1583         unlock_page(page);
1584 out:
1585         return ret;
1586
1587 backout:
1588         spin_unlock(&mm->page_table_lock);
1589         unlock_page(page);
1590         put_page(page);
1591         goto out;
1592 }
1593
1594 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1595                         unsigned long address, int write_access)
1596 {
1597         pte_t *ptep;
1598         pte_t entry;
1599         int ret;
1600         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
1601         struct hstate *h = hstate_vma(vma);
1602
1603         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
1604         if (!ptep)
1605                 return VM_FAULT_OOM;
1606
1607         /*
1608          * Serialize hugepage allocation and instantiation, so that we don't
1609          * get spurious allocation failures if two CPUs race to instantiate
1610          * the same page in the page cache.
1611          */
1612         mutex_lock(&hugetlb_instantiation_mutex);
1613         entry = huge_ptep_get(ptep);
1614         if (huge_pte_none(entry)) {
1615                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1616                 mutex_unlock(&hugetlb_instantiation_mutex);
1617                 return ret;
1618         }
1619
1620         ret = 0;
1621
1622         spin_lock(&mm->page_table_lock);
1623         /* Check for a racing update before calling hugetlb_cow */
1624         if (likely(pte_same(entry, huge_ptep_get(ptep))))
1625                 if (write_access && !pte_write(entry)) {
1626                         struct page *page;
1627                         page = hugetlbfs_pagecache_page(h, vma, address);
1628                         ret = hugetlb_cow(mm, vma, address, ptep, entry, page);
1629                         if (page) {
1630                                 unlock_page(page);
1631                                 put_page(page);
1632                         }
1633                 }
1634         spin_unlock(&mm->page_table_lock);
1635         mutex_unlock(&hugetlb_instantiation_mutex);
1636
1637         return ret;
1638 }
1639
1640 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
1641                         struct page **pages, struct vm_area_struct **vmas,
1642                         unsigned long *position, int *length, int i,
1643                         int write)
1644 {
1645         unsigned long pfn_offset;
1646         unsigned long vaddr = *position;
1647         int remainder = *length;
1648         struct hstate *h = hstate_vma(vma);
1649
1650         spin_lock(&mm->page_table_lock);
1651         while (vaddr < vma->vm_end && remainder) {
1652                 pte_t *pte;
1653                 struct page *page;
1654
1655                 /*
1656                  * Some archs (sparc64, sh*) have multiple pte_ts to
1657                  * each hugepage.  We have to make * sure we get the
1658                  * first, for the page indexing below to work.
1659                  */
1660                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
1661
1662                 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
1663                     (write && !pte_write(huge_ptep_get(pte)))) {
1664                         int ret;
1665
1666                         spin_unlock(&mm->page_table_lock);
1667                         ret = hugetlb_fault(mm, vma, vaddr, write);
1668                         spin_lock(&mm->page_table_lock);
1669                         if (!(ret & VM_FAULT_ERROR))
1670                                 continue;
1671
1672                         remainder = 0;
1673                         if (!i)
1674                                 i = -EFAULT;
1675                         break;
1676                 }
1677
1678                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
1679                 page = pte_page(huge_ptep_get(pte));
1680 same_page:
1681                 if (pages) {
1682                         get_page(page);
1683                         pages[i] = page + pfn_offset;
1684                 }
1685
1686                 if (vmas)
1687                         vmas[i] = vma;
1688
1689                 vaddr += PAGE_SIZE;
1690                 ++pfn_offset;
1691                 --remainder;
1692                 ++i;
1693                 if (vaddr < vma->vm_end && remainder &&
1694                                 pfn_offset < pages_per_huge_page(h)) {
1695                         /*
1696                          * We use pfn_offset to avoid touching the pageframes
1697                          * of this compound page.
1698                          */
1699                         goto same_page;
1700                 }
1701         }
1702         spin_unlock(&mm->page_table_lock);
1703         *length = remainder;
1704         *position = vaddr;
1705
1706         return i;
1707 }
1708
1709 void hugetlb_change_protection(struct vm_area_struct *vma,
1710                 unsigned long address, unsigned long end, pgprot_t newprot)
1711 {
1712         struct mm_struct *mm = vma->vm_mm;
1713         unsigned long start = address;
1714         pte_t *ptep;
1715         pte_t pte;
1716         struct hstate *h = hstate_vma(vma);
1717
1718         BUG_ON(address >= end);
1719         flush_cache_range(vma, address, end);
1720
1721         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1722         spin_lock(&mm->page_table_lock);
1723         for (; address < end; address += huge_page_size(h)) {
1724                 ptep = huge_pte_offset(mm, address);
1725                 if (!ptep)
1726                         continue;
1727                 if (huge_pmd_unshare(mm, &address, ptep))
1728                         continue;
1729                 if (!huge_pte_none(huge_ptep_get(ptep))) {
1730                         pte = huge_ptep_get_and_clear(mm, address, ptep);
1731                         pte = pte_mkhuge(pte_modify(pte, newprot));
1732                         set_huge_pte_at(mm, address, ptep, pte);
1733                 }
1734         }
1735         spin_unlock(&mm->page_table_lock);
1736         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1737
1738         flush_tlb_range(vma, start, end);
1739 }
1740
1741 int hugetlb_reserve_pages(struct inode *inode,
1742                                         long from, long to,
1743                                         struct vm_area_struct *vma)
1744 {
1745         long ret, chg;
1746         struct hstate *h = hstate_inode(inode);
1747
1748         if (vma && vma->vm_flags & VM_NORESERVE)
1749                 return 0;
1750
1751         /*
1752          * Shared mappings base their reservation on the number of pages that
1753          * are already allocated on behalf of the file. Private mappings need
1754          * to reserve the full area even if read-only as mprotect() may be
1755          * called to make the mapping read-write. Assume !vma is a shm mapping
1756          */
1757         if (!vma || vma->vm_flags & VM_SHARED)
1758                 chg = region_chg(&inode->i_mapping->private_list, from, to);
1759         else {
1760                 struct resv_map *resv_map = resv_map_alloc();
1761                 if (!resv_map)
1762                         return -ENOMEM;
1763
1764                 chg = to - from;
1765
1766                 set_vma_resv_map(vma, resv_map);
1767                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
1768         }
1769
1770         if (chg < 0)
1771                 return chg;
1772
1773         if (hugetlb_get_quota(inode->i_mapping, chg))
1774                 return -ENOSPC;
1775         ret = hugetlb_acct_memory(h, chg);
1776         if (ret < 0) {
1777                 hugetlb_put_quota(inode->i_mapping, chg);
1778                 return ret;
1779         }
1780         if (!vma || vma->vm_flags & VM_SHARED)
1781                 region_add(&inode->i_mapping->private_list, from, to);
1782         return 0;
1783 }
1784
1785 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1786 {
1787         struct hstate *h = hstate_inode(inode);
1788         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1789
1790         spin_lock(&inode->i_lock);
1791         inode->i_blocks -= blocks_per_huge_page(h);
1792         spin_unlock(&inode->i_lock);
1793
1794         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1795         hugetlb_acct_memory(h, -(chg - freed));
1796 }