hugetlbfs: common code update for s390
[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 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
26 static unsigned long surplus_huge_pages;
27 static unsigned long nr_overcommit_huge_pages;
28 unsigned long max_huge_pages;
29 unsigned long sysctl_overcommit_huge_pages;
30 static struct list_head hugepage_freelists[MAX_NUMNODES];
31 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
32 static unsigned int free_huge_pages_node[MAX_NUMNODES];
33 static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
36 static int hugetlb_next_nid;
37
38 /*
39  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
40  */
41 static DEFINE_SPINLOCK(hugetlb_lock);
42
43 static void clear_huge_page(struct page *page, unsigned long addr)
44 {
45         int i;
46
47         might_sleep();
48         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
49                 cond_resched();
50                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
51         }
52 }
53
54 static void copy_huge_page(struct page *dst, struct page *src,
55                            unsigned long addr, struct vm_area_struct *vma)
56 {
57         int i;
58
59         might_sleep();
60         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
61                 cond_resched();
62                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
63         }
64 }
65
66 static void enqueue_huge_page(struct page *page)
67 {
68         int nid = page_to_nid(page);
69         list_add(&page->lru, &hugepage_freelists[nid]);
70         free_huge_pages++;
71         free_huge_pages_node[nid]++;
72 }
73
74 static struct page *dequeue_huge_page(void)
75 {
76         int nid;
77         struct page *page = NULL;
78
79         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
80                 if (!list_empty(&hugepage_freelists[nid])) {
81                         page = list_entry(hugepage_freelists[nid].next,
82                                           struct page, lru);
83                         list_del(&page->lru);
84                         free_huge_pages--;
85                         free_huge_pages_node[nid]--;
86                         break;
87                 }
88         }
89         return page;
90 }
91
92 static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
93                                 unsigned long address)
94 {
95         int nid;
96         struct page *page = NULL;
97         struct mempolicy *mpol;
98         nodemask_t *nodemask;
99         struct zonelist *zonelist = huge_zonelist(vma, address,
100                                         htlb_alloc_mask, &mpol, &nodemask);
101         struct zone *zone;
102         struct zoneref *z;
103
104         for_each_zone_zonelist_nodemask(zone, z, zonelist,
105                                                 MAX_NR_ZONES - 1, nodemask) {
106                 nid = zone_to_nid(zone);
107                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
108                     !list_empty(&hugepage_freelists[nid])) {
109                         page = list_entry(hugepage_freelists[nid].next,
110                                           struct page, lru);
111                         list_del(&page->lru);
112                         free_huge_pages--;
113                         free_huge_pages_node[nid]--;
114                         if (vma && vma->vm_flags & VM_MAYSHARE)
115                                 resv_huge_pages--;
116                         break;
117                 }
118         }
119         mpol_cond_put(mpol);
120         return page;
121 }
122
123 static void update_and_free_page(struct page *page)
124 {
125         int i;
126         nr_huge_pages--;
127         nr_huge_pages_node[page_to_nid(page)]--;
128         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
129                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
130                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
131                                 1 << PG_private | 1<< PG_writeback);
132         }
133         set_compound_page_dtor(page, NULL);
134         set_page_refcounted(page);
135         arch_release_hugepage(page);
136         __free_pages(page, HUGETLB_PAGE_ORDER);
137 }
138
139 static void free_huge_page(struct page *page)
140 {
141         int nid = page_to_nid(page);
142         struct address_space *mapping;
143
144         mapping = (struct address_space *) page_private(page);
145         set_page_private(page, 0);
146         BUG_ON(page_count(page));
147         INIT_LIST_HEAD(&page->lru);
148
149         spin_lock(&hugetlb_lock);
150         if (surplus_huge_pages_node[nid]) {
151                 update_and_free_page(page);
152                 surplus_huge_pages--;
153                 surplus_huge_pages_node[nid]--;
154         } else {
155                 enqueue_huge_page(page);
156         }
157         spin_unlock(&hugetlb_lock);
158         if (mapping)
159                 hugetlb_put_quota(mapping, 1);
160 }
161
162 /*
163  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
164  * balanced by operating on them in a round-robin fashion.
165  * Returns 1 if an adjustment was made.
166  */
167 static int adjust_pool_surplus(int delta)
168 {
169         static int prev_nid;
170         int nid = prev_nid;
171         int ret = 0;
172
173         VM_BUG_ON(delta != -1 && delta != 1);
174         do {
175                 nid = next_node(nid, node_online_map);
176                 if (nid == MAX_NUMNODES)
177                         nid = first_node(node_online_map);
178
179                 /* To shrink on this node, there must be a surplus page */
180                 if (delta < 0 && !surplus_huge_pages_node[nid])
181                         continue;
182                 /* Surplus cannot exceed the total number of pages */
183                 if (delta > 0 && surplus_huge_pages_node[nid] >=
184                                                 nr_huge_pages_node[nid])
185                         continue;
186
187                 surplus_huge_pages += delta;
188                 surplus_huge_pages_node[nid] += delta;
189                 ret = 1;
190                 break;
191         } while (nid != prev_nid);
192
193         prev_nid = nid;
194         return ret;
195 }
196
197 static struct page *alloc_fresh_huge_page_node(int nid)
198 {
199         struct page *page;
200
201         page = alloc_pages_node(nid,
202                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
203                 HUGETLB_PAGE_ORDER);
204         if (page) {
205                 if (arch_prepare_hugepage(page)) {
206                         __free_pages(page, HUGETLB_PAGE_ORDER);
207                         return 0;
208                 }
209                 set_compound_page_dtor(page, free_huge_page);
210                 spin_lock(&hugetlb_lock);
211                 nr_huge_pages++;
212                 nr_huge_pages_node[nid]++;
213                 spin_unlock(&hugetlb_lock);
214                 put_page(page); /* free it into the hugepage allocator */
215         }
216
217         return page;
218 }
219
220 static int alloc_fresh_huge_page(void)
221 {
222         struct page *page;
223         int start_nid;
224         int next_nid;
225         int ret = 0;
226
227         start_nid = hugetlb_next_nid;
228
229         do {
230                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
231                 if (page)
232                         ret = 1;
233                 /*
234                  * Use a helper variable to find the next node and then
235                  * copy it back to hugetlb_next_nid afterwards:
236                  * otherwise there's a window in which a racer might
237                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
238                  * But we don't need to use a spin_lock here: it really
239                  * doesn't matter if occasionally a racer chooses the
240                  * same nid as we do.  Move nid forward in the mask even
241                  * if we just successfully allocated a hugepage so that
242                  * the next caller gets hugepages on the next node.
243                  */
244                 next_nid = next_node(hugetlb_next_nid, node_online_map);
245                 if (next_nid == MAX_NUMNODES)
246                         next_nid = first_node(node_online_map);
247                 hugetlb_next_nid = next_nid;
248         } while (!page && hugetlb_next_nid != start_nid);
249
250         if (ret)
251                 count_vm_event(HTLB_BUDDY_PGALLOC);
252         else
253                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
254
255         return ret;
256 }
257
258 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
259                                                 unsigned long address)
260 {
261         struct page *page;
262         unsigned int nid;
263
264         /*
265          * Assume we will successfully allocate the surplus page to
266          * prevent racing processes from causing the surplus to exceed
267          * overcommit
268          *
269          * This however introduces a different race, where a process B
270          * tries to grow the static hugepage pool while alloc_pages() is
271          * called by process A. B will only examine the per-node
272          * counters in determining if surplus huge pages can be
273          * converted to normal huge pages in adjust_pool_surplus(). A
274          * won't be able to increment the per-node counter, until the
275          * lock is dropped by B, but B doesn't drop hugetlb_lock until
276          * no more huge pages can be converted from surplus to normal
277          * state (and doesn't try to convert again). Thus, we have a
278          * case where a surplus huge page exists, the pool is grown, and
279          * the surplus huge page still exists after, even though it
280          * should just have been converted to a normal huge page. This
281          * does not leak memory, though, as the hugepage will be freed
282          * once it is out of use. It also does not allow the counters to
283          * go out of whack in adjust_pool_surplus() as we don't modify
284          * the node values until we've gotten the hugepage and only the
285          * per-node value is checked there.
286          */
287         spin_lock(&hugetlb_lock);
288         if (surplus_huge_pages >= nr_overcommit_huge_pages) {
289                 spin_unlock(&hugetlb_lock);
290                 return NULL;
291         } else {
292                 nr_huge_pages++;
293                 surplus_huge_pages++;
294         }
295         spin_unlock(&hugetlb_lock);
296
297         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
298                                         HUGETLB_PAGE_ORDER);
299
300         spin_lock(&hugetlb_lock);
301         if (page) {
302                 /*
303                  * This page is now managed by the hugetlb allocator and has
304                  * no users -- drop the buddy allocator's reference.
305                  */
306                 put_page_testzero(page);
307                 VM_BUG_ON(page_count(page));
308                 nid = page_to_nid(page);
309                 set_compound_page_dtor(page, free_huge_page);
310                 /*
311                  * We incremented the global counters already
312                  */
313                 nr_huge_pages_node[nid]++;
314                 surplus_huge_pages_node[nid]++;
315                 __count_vm_event(HTLB_BUDDY_PGALLOC);
316         } else {
317                 nr_huge_pages--;
318                 surplus_huge_pages--;
319                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
320         }
321         spin_unlock(&hugetlb_lock);
322
323         return page;
324 }
325
326 /*
327  * Increase the hugetlb pool such that it can accomodate a reservation
328  * of size 'delta'.
329  */
330 static int gather_surplus_pages(int delta)
331 {
332         struct list_head surplus_list;
333         struct page *page, *tmp;
334         int ret, i;
335         int needed, allocated;
336
337         needed = (resv_huge_pages + delta) - free_huge_pages;
338         if (needed <= 0) {
339                 resv_huge_pages += delta;
340                 return 0;
341         }
342
343         allocated = 0;
344         INIT_LIST_HEAD(&surplus_list);
345
346         ret = -ENOMEM;
347 retry:
348         spin_unlock(&hugetlb_lock);
349         for (i = 0; i < needed; i++) {
350                 page = alloc_buddy_huge_page(NULL, 0);
351                 if (!page) {
352                         /*
353                          * We were not able to allocate enough pages to
354                          * satisfy the entire reservation so we free what
355                          * we've allocated so far.
356                          */
357                         spin_lock(&hugetlb_lock);
358                         needed = 0;
359                         goto free;
360                 }
361
362                 list_add(&page->lru, &surplus_list);
363         }
364         allocated += needed;
365
366         /*
367          * After retaking hugetlb_lock, we need to recalculate 'needed'
368          * because either resv_huge_pages or free_huge_pages may have changed.
369          */
370         spin_lock(&hugetlb_lock);
371         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
372         if (needed > 0)
373                 goto retry;
374
375         /*
376          * The surplus_list now contains _at_least_ the number of extra pages
377          * needed to accomodate the reservation.  Add the appropriate number
378          * of pages to the hugetlb pool and free the extras back to the buddy
379          * allocator.  Commit the entire reservation here to prevent another
380          * process from stealing the pages as they are added to the pool but
381          * before they are reserved.
382          */
383         needed += allocated;
384         resv_huge_pages += delta;
385         ret = 0;
386 free:
387         /* Free the needed pages to the hugetlb pool */
388         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
389                 if ((--needed) < 0)
390                         break;
391                 list_del(&page->lru);
392                 enqueue_huge_page(page);
393         }
394
395         /* Free unnecessary surplus pages to the buddy allocator */
396         if (!list_empty(&surplus_list)) {
397                 spin_unlock(&hugetlb_lock);
398                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
399                         list_del(&page->lru);
400                         /*
401                          * The page has a reference count of zero already, so
402                          * call free_huge_page directly instead of using
403                          * put_page.  This must be done with hugetlb_lock
404                          * unlocked which is safe because free_huge_page takes
405                          * hugetlb_lock before deciding how to free the page.
406                          */
407                         free_huge_page(page);
408                 }
409                 spin_lock(&hugetlb_lock);
410         }
411
412         return ret;
413 }
414
415 /*
416  * When releasing a hugetlb pool reservation, any surplus pages that were
417  * allocated to satisfy the reservation must be explicitly freed if they were
418  * never used.
419  */
420 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
421 {
422         static int nid = -1;
423         struct page *page;
424         unsigned long nr_pages;
425
426         /*
427          * We want to release as many surplus pages as possible, spread
428          * evenly across all nodes. Iterate across all nodes until we
429          * can no longer free unreserved surplus pages. This occurs when
430          * the nodes with surplus pages have no free pages.
431          */
432         unsigned long remaining_iterations = num_online_nodes();
433
434         /* Uncommit the reservation */
435         resv_huge_pages -= unused_resv_pages;
436
437         nr_pages = min(unused_resv_pages, surplus_huge_pages);
438
439         while (remaining_iterations-- && nr_pages) {
440                 nid = next_node(nid, node_online_map);
441                 if (nid == MAX_NUMNODES)
442                         nid = first_node(node_online_map);
443
444                 if (!surplus_huge_pages_node[nid])
445                         continue;
446
447                 if (!list_empty(&hugepage_freelists[nid])) {
448                         page = list_entry(hugepage_freelists[nid].next,
449                                           struct page, lru);
450                         list_del(&page->lru);
451                         update_and_free_page(page);
452                         free_huge_pages--;
453                         free_huge_pages_node[nid]--;
454                         surplus_huge_pages--;
455                         surplus_huge_pages_node[nid]--;
456                         nr_pages--;
457                         remaining_iterations = num_online_nodes();
458                 }
459         }
460 }
461
462
463 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
464                                                 unsigned long addr)
465 {
466         struct page *page;
467
468         spin_lock(&hugetlb_lock);
469         page = dequeue_huge_page_vma(vma, addr);
470         spin_unlock(&hugetlb_lock);
471         return page ? page : ERR_PTR(-VM_FAULT_OOM);
472 }
473
474 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
475                                                 unsigned long addr)
476 {
477         struct page *page = NULL;
478
479         if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
480                 return ERR_PTR(-VM_FAULT_SIGBUS);
481
482         spin_lock(&hugetlb_lock);
483         if (free_huge_pages > resv_huge_pages)
484                 page = dequeue_huge_page_vma(vma, addr);
485         spin_unlock(&hugetlb_lock);
486         if (!page) {
487                 page = alloc_buddy_huge_page(vma, addr);
488                 if (!page) {
489                         hugetlb_put_quota(vma->vm_file->f_mapping, 1);
490                         return ERR_PTR(-VM_FAULT_OOM);
491                 }
492         }
493         return page;
494 }
495
496 static struct page *alloc_huge_page(struct vm_area_struct *vma,
497                                     unsigned long addr)
498 {
499         struct page *page;
500         struct address_space *mapping = vma->vm_file->f_mapping;
501
502         if (vma->vm_flags & VM_MAYSHARE)
503                 page = alloc_huge_page_shared(vma, addr);
504         else
505                 page = alloc_huge_page_private(vma, addr);
506
507         if (!IS_ERR(page)) {
508                 set_page_refcounted(page);
509                 set_page_private(page, (unsigned long) mapping);
510         }
511         return page;
512 }
513
514 static int __init hugetlb_init(void)
515 {
516         unsigned long i;
517
518         if (HPAGE_SHIFT == 0)
519                 return 0;
520
521         for (i = 0; i < MAX_NUMNODES; ++i)
522                 INIT_LIST_HEAD(&hugepage_freelists[i]);
523
524         hugetlb_next_nid = first_node(node_online_map);
525
526         for (i = 0; i < max_huge_pages; ++i) {
527                 if (!alloc_fresh_huge_page())
528                         break;
529         }
530         max_huge_pages = free_huge_pages = nr_huge_pages = i;
531         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
532         return 0;
533 }
534 module_init(hugetlb_init);
535
536 static int __init hugetlb_setup(char *s)
537 {
538         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
539                 max_huge_pages = 0;
540         return 1;
541 }
542 __setup("hugepages=", hugetlb_setup);
543
544 static unsigned int cpuset_mems_nr(unsigned int *array)
545 {
546         int node;
547         unsigned int nr = 0;
548
549         for_each_node_mask(node, cpuset_current_mems_allowed)
550                 nr += array[node];
551
552         return nr;
553 }
554
555 #ifdef CONFIG_SYSCTL
556 #ifdef CONFIG_HIGHMEM
557 static void try_to_free_low(unsigned long count)
558 {
559         int i;
560
561         for (i = 0; i < MAX_NUMNODES; ++i) {
562                 struct page *page, *next;
563                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
564                         if (count >= nr_huge_pages)
565                                 return;
566                         if (PageHighMem(page))
567                                 continue;
568                         list_del(&page->lru);
569                         update_and_free_page(page);
570                         free_huge_pages--;
571                         free_huge_pages_node[page_to_nid(page)]--;
572                 }
573         }
574 }
575 #else
576 static inline void try_to_free_low(unsigned long count)
577 {
578 }
579 #endif
580
581 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
582 static unsigned long set_max_huge_pages(unsigned long count)
583 {
584         unsigned long min_count, ret;
585
586         /*
587          * Increase the pool size
588          * First take pages out of surplus state.  Then make up the
589          * remaining difference by allocating fresh huge pages.
590          *
591          * We might race with alloc_buddy_huge_page() here and be unable
592          * to convert a surplus huge page to a normal huge page. That is
593          * not critical, though, it just means the overall size of the
594          * pool might be one hugepage larger than it needs to be, but
595          * within all the constraints specified by the sysctls.
596          */
597         spin_lock(&hugetlb_lock);
598         while (surplus_huge_pages && count > persistent_huge_pages) {
599                 if (!adjust_pool_surplus(-1))
600                         break;
601         }
602
603         while (count > persistent_huge_pages) {
604                 int ret;
605                 /*
606                  * If this allocation races such that we no longer need the
607                  * page, free_huge_page will handle it by freeing the page
608                  * and reducing the surplus.
609                  */
610                 spin_unlock(&hugetlb_lock);
611                 ret = alloc_fresh_huge_page();
612                 spin_lock(&hugetlb_lock);
613                 if (!ret)
614                         goto out;
615
616         }
617
618         /*
619          * Decrease the pool size
620          * First return free pages to the buddy allocator (being careful
621          * to keep enough around to satisfy reservations).  Then place
622          * pages into surplus state as needed so the pool will shrink
623          * to the desired size as pages become free.
624          *
625          * By placing pages into the surplus state independent of the
626          * overcommit value, we are allowing the surplus pool size to
627          * exceed overcommit. There are few sane options here. Since
628          * alloc_buddy_huge_page() is checking the global counter,
629          * though, we'll note that we're not allowed to exceed surplus
630          * and won't grow the pool anywhere else. Not until one of the
631          * sysctls are changed, or the surplus pages go out of use.
632          */
633         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
634         min_count = max(count, min_count);
635         try_to_free_low(min_count);
636         while (min_count < persistent_huge_pages) {
637                 struct page *page = dequeue_huge_page();
638                 if (!page)
639                         break;
640                 update_and_free_page(page);
641         }
642         while (count < persistent_huge_pages) {
643                 if (!adjust_pool_surplus(1))
644                         break;
645         }
646 out:
647         ret = persistent_huge_pages;
648         spin_unlock(&hugetlb_lock);
649         return ret;
650 }
651
652 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
653                            struct file *file, void __user *buffer,
654                            size_t *length, loff_t *ppos)
655 {
656         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
657         max_huge_pages = set_max_huge_pages(max_huge_pages);
658         return 0;
659 }
660
661 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
662                         struct file *file, void __user *buffer,
663                         size_t *length, loff_t *ppos)
664 {
665         proc_dointvec(table, write, file, buffer, length, ppos);
666         if (hugepages_treat_as_movable)
667                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
668         else
669                 htlb_alloc_mask = GFP_HIGHUSER;
670         return 0;
671 }
672
673 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
674                         struct file *file, void __user *buffer,
675                         size_t *length, loff_t *ppos)
676 {
677         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
678         spin_lock(&hugetlb_lock);
679         nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
680         spin_unlock(&hugetlb_lock);
681         return 0;
682 }
683
684 #endif /* CONFIG_SYSCTL */
685
686 int hugetlb_report_meminfo(char *buf)
687 {
688         return sprintf(buf,
689                         "HugePages_Total: %5lu\n"
690                         "HugePages_Free:  %5lu\n"
691                         "HugePages_Rsvd:  %5lu\n"
692                         "HugePages_Surp:  %5lu\n"
693                         "Hugepagesize:    %5lu kB\n",
694                         nr_huge_pages,
695                         free_huge_pages,
696                         resv_huge_pages,
697                         surplus_huge_pages,
698                         HPAGE_SIZE/1024);
699 }
700
701 int hugetlb_report_node_meminfo(int nid, char *buf)
702 {
703         return sprintf(buf,
704                 "Node %d HugePages_Total: %5u\n"
705                 "Node %d HugePages_Free:  %5u\n"
706                 "Node %d HugePages_Surp:  %5u\n",
707                 nid, nr_huge_pages_node[nid],
708                 nid, free_huge_pages_node[nid],
709                 nid, surplus_huge_pages_node[nid]);
710 }
711
712 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
713 unsigned long hugetlb_total_pages(void)
714 {
715         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
716 }
717
718 /*
719  * We cannot handle pagefaults against hugetlb pages at all.  They cause
720  * handle_mm_fault() to try to instantiate regular-sized pages in the
721  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
722  * this far.
723  */
724 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
725 {
726         BUG();
727         return 0;
728 }
729
730 struct vm_operations_struct hugetlb_vm_ops = {
731         .fault = hugetlb_vm_op_fault,
732 };
733
734 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
735                                 int writable)
736 {
737         pte_t entry;
738
739         if (writable) {
740                 entry =
741                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
742         } else {
743                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
744         }
745         entry = pte_mkyoung(entry);
746         entry = pte_mkhuge(entry);
747
748         return entry;
749 }
750
751 static void set_huge_ptep_writable(struct vm_area_struct *vma,
752                                    unsigned long address, pte_t *ptep)
753 {
754         pte_t entry;
755
756         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
757         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
758                 update_mmu_cache(vma, address, entry);
759         }
760 }
761
762
763 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
764                             struct vm_area_struct *vma)
765 {
766         pte_t *src_pte, *dst_pte, entry;
767         struct page *ptepage;
768         unsigned long addr;
769         int cow;
770
771         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
772
773         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
774                 src_pte = huge_pte_offset(src, addr);
775                 if (!src_pte)
776                         continue;
777                 dst_pte = huge_pte_alloc(dst, addr);
778                 if (!dst_pte)
779                         goto nomem;
780
781                 /* If the pagetables are shared don't copy or take references */
782                 if (dst_pte == src_pte)
783                         continue;
784
785                 spin_lock(&dst->page_table_lock);
786                 spin_lock(&src->page_table_lock);
787                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
788                         if (cow)
789                                 huge_ptep_set_wrprotect(src, addr, src_pte);
790                         entry = huge_ptep_get(src_pte);
791                         ptepage = pte_page(entry);
792                         get_page(ptepage);
793                         set_huge_pte_at(dst, addr, dst_pte, entry);
794                 }
795                 spin_unlock(&src->page_table_lock);
796                 spin_unlock(&dst->page_table_lock);
797         }
798         return 0;
799
800 nomem:
801         return -ENOMEM;
802 }
803
804 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
805                             unsigned long end)
806 {
807         struct mm_struct *mm = vma->vm_mm;
808         unsigned long address;
809         pte_t *ptep;
810         pte_t pte;
811         struct page *page;
812         struct page *tmp;
813         /*
814          * A page gathering list, protected by per file i_mmap_lock. The
815          * lock is used to avoid list corruption from multiple unmapping
816          * of the same page since we are using page->lru.
817          */
818         LIST_HEAD(page_list);
819
820         WARN_ON(!is_vm_hugetlb_page(vma));
821         BUG_ON(start & ~HPAGE_MASK);
822         BUG_ON(end & ~HPAGE_MASK);
823
824         spin_lock(&mm->page_table_lock);
825         for (address = start; address < end; address += HPAGE_SIZE) {
826                 ptep = huge_pte_offset(mm, address);
827                 if (!ptep)
828                         continue;
829
830                 if (huge_pmd_unshare(mm, &address, ptep))
831                         continue;
832
833                 pte = huge_ptep_get_and_clear(mm, address, ptep);
834                 if (huge_pte_none(pte))
835                         continue;
836
837                 page = pte_page(pte);
838                 if (pte_dirty(pte))
839                         set_page_dirty(page);
840                 list_add(&page->lru, &page_list);
841         }
842         spin_unlock(&mm->page_table_lock);
843         flush_tlb_range(vma, start, end);
844         list_for_each_entry_safe(page, tmp, &page_list, lru) {
845                 list_del(&page->lru);
846                 put_page(page);
847         }
848 }
849
850 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
851                           unsigned long end)
852 {
853         /*
854          * It is undesirable to test vma->vm_file as it should be non-null
855          * for valid hugetlb area. However, vm_file will be NULL in the error
856          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
857          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
858          * to clean up. Since no pte has actually been setup, it is safe to
859          * do nothing in this case.
860          */
861         if (vma->vm_file) {
862                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
863                 __unmap_hugepage_range(vma, start, end);
864                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
865         }
866 }
867
868 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
869                         unsigned long address, pte_t *ptep, pte_t pte)
870 {
871         struct page *old_page, *new_page;
872         int avoidcopy;
873
874         old_page = pte_page(pte);
875
876         /* If no-one else is actually using this page, avoid the copy
877          * and just make the page writable */
878         avoidcopy = (page_count(old_page) == 1);
879         if (avoidcopy) {
880                 set_huge_ptep_writable(vma, address, ptep);
881                 return 0;
882         }
883
884         page_cache_get(old_page);
885         new_page = alloc_huge_page(vma, address);
886
887         if (IS_ERR(new_page)) {
888                 page_cache_release(old_page);
889                 return -PTR_ERR(new_page);
890         }
891
892         spin_unlock(&mm->page_table_lock);
893         copy_huge_page(new_page, old_page, address, vma);
894         __SetPageUptodate(new_page);
895         spin_lock(&mm->page_table_lock);
896
897         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
898         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
899                 /* Break COW */
900                 huge_ptep_clear_flush(vma, address, ptep);
901                 set_huge_pte_at(mm, address, ptep,
902                                 make_huge_pte(vma, new_page, 1));
903                 /* Make the old page be freed below */
904                 new_page = old_page;
905         }
906         page_cache_release(new_page);
907         page_cache_release(old_page);
908         return 0;
909 }
910
911 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
912                         unsigned long address, pte_t *ptep, int write_access)
913 {
914         int ret = VM_FAULT_SIGBUS;
915         unsigned long idx;
916         unsigned long size;
917         struct page *page;
918         struct address_space *mapping;
919         pte_t new_pte;
920
921         mapping = vma->vm_file->f_mapping;
922         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
923                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
924
925         /*
926          * Use page lock to guard against racing truncation
927          * before we get page_table_lock.
928          */
929 retry:
930         page = find_lock_page(mapping, idx);
931         if (!page) {
932                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
933                 if (idx >= size)
934                         goto out;
935                 page = alloc_huge_page(vma, address);
936                 if (IS_ERR(page)) {
937                         ret = -PTR_ERR(page);
938                         goto out;
939                 }
940                 clear_huge_page(page, address);
941                 __SetPageUptodate(page);
942
943                 if (vma->vm_flags & VM_SHARED) {
944                         int err;
945                         struct inode *inode = mapping->host;
946
947                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
948                         if (err) {
949                                 put_page(page);
950                                 if (err == -EEXIST)
951                                         goto retry;
952                                 goto out;
953                         }
954
955                         spin_lock(&inode->i_lock);
956                         inode->i_blocks += BLOCKS_PER_HUGEPAGE;
957                         spin_unlock(&inode->i_lock);
958                 } else
959                         lock_page(page);
960         }
961
962         spin_lock(&mm->page_table_lock);
963         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
964         if (idx >= size)
965                 goto backout;
966
967         ret = 0;
968         if (!huge_pte_none(huge_ptep_get(ptep)))
969                 goto backout;
970
971         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
972                                 && (vma->vm_flags & VM_SHARED)));
973         set_huge_pte_at(mm, address, ptep, new_pte);
974
975         if (write_access && !(vma->vm_flags & VM_SHARED)) {
976                 /* Optimization, do the COW without a second fault */
977                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
978         }
979
980         spin_unlock(&mm->page_table_lock);
981         unlock_page(page);
982 out:
983         return ret;
984
985 backout:
986         spin_unlock(&mm->page_table_lock);
987         unlock_page(page);
988         put_page(page);
989         goto out;
990 }
991
992 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
993                         unsigned long address, int write_access)
994 {
995         pte_t *ptep;
996         pte_t entry;
997         int ret;
998         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
999
1000         ptep = huge_pte_alloc(mm, address);
1001         if (!ptep)
1002                 return VM_FAULT_OOM;
1003
1004         /*
1005          * Serialize hugepage allocation and instantiation, so that we don't
1006          * get spurious allocation failures if two CPUs race to instantiate
1007          * the same page in the page cache.
1008          */
1009         mutex_lock(&hugetlb_instantiation_mutex);
1010         entry = huge_ptep_get(ptep);
1011         if (huge_pte_none(entry)) {
1012                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1013                 mutex_unlock(&hugetlb_instantiation_mutex);
1014                 return ret;
1015         }
1016
1017         ret = 0;
1018
1019         spin_lock(&mm->page_table_lock);
1020         /* Check for a racing update before calling hugetlb_cow */
1021         if (likely(pte_same(entry, huge_ptep_get(ptep))))
1022                 if (write_access && !pte_write(entry))
1023                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
1024         spin_unlock(&mm->page_table_lock);
1025         mutex_unlock(&hugetlb_instantiation_mutex);
1026
1027         return ret;
1028 }
1029
1030 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
1031                         struct page **pages, struct vm_area_struct **vmas,
1032                         unsigned long *position, int *length, int i,
1033                         int write)
1034 {
1035         unsigned long pfn_offset;
1036         unsigned long vaddr = *position;
1037         int remainder = *length;
1038
1039         spin_lock(&mm->page_table_lock);
1040         while (vaddr < vma->vm_end && remainder) {
1041                 pte_t *pte;
1042                 struct page *page;
1043
1044                 /*
1045                  * Some archs (sparc64, sh*) have multiple pte_ts to
1046                  * each hugepage.  We have to make * sure we get the
1047                  * first, for the page indexing below to work.
1048                  */
1049                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1050
1051                 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
1052                     (write && !pte_write(huge_ptep_get(pte)))) {
1053                         int ret;
1054
1055                         spin_unlock(&mm->page_table_lock);
1056                         ret = hugetlb_fault(mm, vma, vaddr, write);
1057                         spin_lock(&mm->page_table_lock);
1058                         if (!(ret & VM_FAULT_ERROR))
1059                                 continue;
1060
1061                         remainder = 0;
1062                         if (!i)
1063                                 i = -EFAULT;
1064                         break;
1065                 }
1066
1067                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1068                 page = pte_page(huge_ptep_get(pte));
1069 same_page:
1070                 if (pages) {
1071                         get_page(page);
1072                         pages[i] = page + pfn_offset;
1073                 }
1074
1075                 if (vmas)
1076                         vmas[i] = vma;
1077
1078                 vaddr += PAGE_SIZE;
1079                 ++pfn_offset;
1080                 --remainder;
1081                 ++i;
1082                 if (vaddr < vma->vm_end && remainder &&
1083                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1084                         /*
1085                          * We use pfn_offset to avoid touching the pageframes
1086                          * of this compound page.
1087                          */
1088                         goto same_page;
1089                 }
1090         }
1091         spin_unlock(&mm->page_table_lock);
1092         *length = remainder;
1093         *position = vaddr;
1094
1095         return i;
1096 }
1097
1098 void hugetlb_change_protection(struct vm_area_struct *vma,
1099                 unsigned long address, unsigned long end, pgprot_t newprot)
1100 {
1101         struct mm_struct *mm = vma->vm_mm;
1102         unsigned long start = address;
1103         pte_t *ptep;
1104         pte_t pte;
1105
1106         BUG_ON(address >= end);
1107         flush_cache_range(vma, address, end);
1108
1109         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1110         spin_lock(&mm->page_table_lock);
1111         for (; address < end; address += HPAGE_SIZE) {
1112                 ptep = huge_pte_offset(mm, address);
1113                 if (!ptep)
1114                         continue;
1115                 if (huge_pmd_unshare(mm, &address, ptep))
1116                         continue;
1117                 if (!huge_pte_none(huge_ptep_get(ptep))) {
1118                         pte = huge_ptep_get_and_clear(mm, address, ptep);
1119                         pte = pte_mkhuge(pte_modify(pte, newprot));
1120                         set_huge_pte_at(mm, address, ptep, pte);
1121                 }
1122         }
1123         spin_unlock(&mm->page_table_lock);
1124         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1125
1126         flush_tlb_range(vma, start, end);
1127 }
1128
1129 struct file_region {
1130         struct list_head link;
1131         long from;
1132         long to;
1133 };
1134
1135 static long region_add(struct list_head *head, long f, long t)
1136 {
1137         struct file_region *rg, *nrg, *trg;
1138
1139         /* Locate the region we are either in or before. */
1140         list_for_each_entry(rg, head, link)
1141                 if (f <= rg->to)
1142                         break;
1143
1144         /* Round our left edge to the current segment if it encloses us. */
1145         if (f > rg->from)
1146                 f = rg->from;
1147
1148         /* Check for and consume any regions we now overlap with. */
1149         nrg = rg;
1150         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1151                 if (&rg->link == head)
1152                         break;
1153                 if (rg->from > t)
1154                         break;
1155
1156                 /* If this area reaches higher then extend our area to
1157                  * include it completely.  If this is not the first area
1158                  * which we intend to reuse, free it. */
1159                 if (rg->to > t)
1160                         t = rg->to;
1161                 if (rg != nrg) {
1162                         list_del(&rg->link);
1163                         kfree(rg);
1164                 }
1165         }
1166         nrg->from = f;
1167         nrg->to = t;
1168         return 0;
1169 }
1170
1171 static long region_chg(struct list_head *head, long f, long t)
1172 {
1173         struct file_region *rg, *nrg;
1174         long chg = 0;
1175
1176         /* Locate the region we are before or in. */
1177         list_for_each_entry(rg, head, link)
1178                 if (f <= rg->to)
1179                         break;
1180
1181         /* If we are below the current region then a new region is required.
1182          * Subtle, allocate a new region at the position but make it zero
1183          * size such that we can guarantee to record the reservation. */
1184         if (&rg->link == head || t < rg->from) {
1185                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1186                 if (!nrg)
1187                         return -ENOMEM;
1188                 nrg->from = f;
1189                 nrg->to   = f;
1190                 INIT_LIST_HEAD(&nrg->link);
1191                 list_add(&nrg->link, rg->link.prev);
1192
1193                 return t - f;
1194         }
1195
1196         /* Round our left edge to the current segment if it encloses us. */
1197         if (f > rg->from)
1198                 f = rg->from;
1199         chg = t - f;
1200
1201         /* Check for and consume any regions we now overlap with. */
1202         list_for_each_entry(rg, rg->link.prev, link) {
1203                 if (&rg->link == head)
1204                         break;
1205                 if (rg->from > t)
1206                         return chg;
1207
1208                 /* We overlap with this area, if it extends futher than
1209                  * us then we must extend ourselves.  Account for its
1210                  * existing reservation. */
1211                 if (rg->to > t) {
1212                         chg += rg->to - t;
1213                         t = rg->to;
1214                 }
1215                 chg -= rg->to - rg->from;
1216         }
1217         return chg;
1218 }
1219
1220 static long region_truncate(struct list_head *head, long end)
1221 {
1222         struct file_region *rg, *trg;
1223         long chg = 0;
1224
1225         /* Locate the region we are either in or before. */
1226         list_for_each_entry(rg, head, link)
1227                 if (end <= rg->to)
1228                         break;
1229         if (&rg->link == head)
1230                 return 0;
1231
1232         /* If we are in the middle of a region then adjust it. */
1233         if (end > rg->from) {
1234                 chg = rg->to - end;
1235                 rg->to = end;
1236                 rg = list_entry(rg->link.next, typeof(*rg), link);
1237         }
1238
1239         /* Drop any remaining regions. */
1240         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1241                 if (&rg->link == head)
1242                         break;
1243                 chg += rg->to - rg->from;
1244                 list_del(&rg->link);
1245                 kfree(rg);
1246         }
1247         return chg;
1248 }
1249
1250 static int hugetlb_acct_memory(long delta)
1251 {
1252         int ret = -ENOMEM;
1253
1254         spin_lock(&hugetlb_lock);
1255         /*
1256          * When cpuset is configured, it breaks the strict hugetlb page
1257          * reservation as the accounting is done on a global variable. Such
1258          * reservation is completely rubbish in the presence of cpuset because
1259          * the reservation is not checked against page availability for the
1260          * current cpuset. Application can still potentially OOM'ed by kernel
1261          * with lack of free htlb page in cpuset that the task is in.
1262          * Attempt to enforce strict accounting with cpuset is almost
1263          * impossible (or too ugly) because cpuset is too fluid that
1264          * task or memory node can be dynamically moved between cpusets.
1265          *
1266          * The change of semantics for shared hugetlb mapping with cpuset is
1267          * undesirable. However, in order to preserve some of the semantics,
1268          * we fall back to check against current free page availability as
1269          * a best attempt and hopefully to minimize the impact of changing
1270          * semantics that cpuset has.
1271          */
1272         if (delta > 0) {
1273                 if (gather_surplus_pages(delta) < 0)
1274                         goto out;
1275
1276                 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1277                         return_unused_surplus_pages(delta);
1278                         goto out;
1279                 }
1280         }
1281
1282         ret = 0;
1283         if (delta < 0)
1284                 return_unused_surplus_pages((unsigned long) -delta);
1285
1286 out:
1287         spin_unlock(&hugetlb_lock);
1288         return ret;
1289 }
1290
1291 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1292 {
1293         long ret, chg;
1294
1295         chg = region_chg(&inode->i_mapping->private_list, from, to);
1296         if (chg < 0)
1297                 return chg;
1298
1299         if (hugetlb_get_quota(inode->i_mapping, chg))
1300                 return -ENOSPC;
1301         ret = hugetlb_acct_memory(chg);
1302         if (ret < 0) {
1303                 hugetlb_put_quota(inode->i_mapping, chg);
1304                 return ret;
1305         }
1306         region_add(&inode->i_mapping->private_list, from, to);
1307         return 0;
1308 }
1309
1310 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1311 {
1312         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1313
1314         spin_lock(&inode->i_lock);
1315         inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1316         spin_unlock(&inode->i_lock);
1317
1318         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1319         hugetlb_acct_memory(-(chg - freed));
1320 }