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