2c5c9ee4220de4e9bb06addd6137ac0ce243d273
[sfrench/cifs-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|
203                                                 __GFP_REPEAT|__GFP_NOWARN,
204                 HUGETLB_PAGE_ORDER);
205         if (page) {
206                 if (arch_prepare_hugepage(page)) {
207                         __free_pages(page, HUGETLB_PAGE_ORDER);
208                         return NULL;
209                 }
210                 set_compound_page_dtor(page, free_huge_page);
211                 spin_lock(&hugetlb_lock);
212                 nr_huge_pages++;
213                 nr_huge_pages_node[nid]++;
214                 spin_unlock(&hugetlb_lock);
215                 put_page(page); /* free it into the hugepage allocator */
216         }
217
218         return page;
219 }
220
221 static int alloc_fresh_huge_page(void)
222 {
223         struct page *page;
224         int start_nid;
225         int next_nid;
226         int ret = 0;
227
228         start_nid = hugetlb_next_nid;
229
230         do {
231                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
232                 if (page)
233                         ret = 1;
234                 /*
235                  * Use a helper variable to find the next node and then
236                  * copy it back to hugetlb_next_nid afterwards:
237                  * otherwise there's a window in which a racer might
238                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
239                  * But we don't need to use a spin_lock here: it really
240                  * doesn't matter if occasionally a racer chooses the
241                  * same nid as we do.  Move nid forward in the mask even
242                  * if we just successfully allocated a hugepage so that
243                  * the next caller gets hugepages on the next node.
244                  */
245                 next_nid = next_node(hugetlb_next_nid, node_online_map);
246                 if (next_nid == MAX_NUMNODES)
247                         next_nid = first_node(node_online_map);
248                 hugetlb_next_nid = next_nid;
249         } while (!page && hugetlb_next_nid != start_nid);
250
251         if (ret)
252                 count_vm_event(HTLB_BUDDY_PGALLOC);
253         else
254                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
255
256         return ret;
257 }
258
259 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
260                                                 unsigned long address)
261 {
262         struct page *page;
263         unsigned int nid;
264
265         /*
266          * Assume we will successfully allocate the surplus page to
267          * prevent racing processes from causing the surplus to exceed
268          * overcommit
269          *
270          * This however introduces a different race, where a process B
271          * tries to grow the static hugepage pool while alloc_pages() is
272          * called by process A. B will only examine the per-node
273          * counters in determining if surplus huge pages can be
274          * converted to normal huge pages in adjust_pool_surplus(). A
275          * won't be able to increment the per-node counter, until the
276          * lock is dropped by B, but B doesn't drop hugetlb_lock until
277          * no more huge pages can be converted from surplus to normal
278          * state (and doesn't try to convert again). Thus, we have a
279          * case where a surplus huge page exists, the pool is grown, and
280          * the surplus huge page still exists after, even though it
281          * should just have been converted to a normal huge page. This
282          * does not leak memory, though, as the hugepage will be freed
283          * once it is out of use. It also does not allow the counters to
284          * go out of whack in adjust_pool_surplus() as we don't modify
285          * the node values until we've gotten the hugepage and only the
286          * per-node value is checked there.
287          */
288         spin_lock(&hugetlb_lock);
289         if (surplus_huge_pages >= nr_overcommit_huge_pages) {
290                 spin_unlock(&hugetlb_lock);
291                 return NULL;
292         } else {
293                 nr_huge_pages++;
294                 surplus_huge_pages++;
295         }
296         spin_unlock(&hugetlb_lock);
297
298         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
299                                         __GFP_REPEAT|__GFP_NOWARN,
300                                         HUGETLB_PAGE_ORDER);
301
302         spin_lock(&hugetlb_lock);
303         if (page) {
304                 /*
305                  * This page is now managed by the hugetlb allocator and has
306                  * no users -- drop the buddy allocator's reference.
307                  */
308                 put_page_testzero(page);
309                 VM_BUG_ON(page_count(page));
310                 nid = page_to_nid(page);
311                 set_compound_page_dtor(page, free_huge_page);
312                 /*
313                  * We incremented the global counters already
314                  */
315                 nr_huge_pages_node[nid]++;
316                 surplus_huge_pages_node[nid]++;
317                 __count_vm_event(HTLB_BUDDY_PGALLOC);
318         } else {
319                 nr_huge_pages--;
320                 surplus_huge_pages--;
321                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
322         }
323         spin_unlock(&hugetlb_lock);
324
325         return page;
326 }
327
328 /*
329  * Increase the hugetlb pool such that it can accomodate a reservation
330  * of size 'delta'.
331  */
332 static int gather_surplus_pages(int delta)
333 {
334         struct list_head surplus_list;
335         struct page *page, *tmp;
336         int ret, i;
337         int needed, allocated;
338
339         needed = (resv_huge_pages + delta) - free_huge_pages;
340         if (needed <= 0) {
341                 resv_huge_pages += delta;
342                 return 0;
343         }
344
345         allocated = 0;
346         INIT_LIST_HEAD(&surplus_list);
347
348         ret = -ENOMEM;
349 retry:
350         spin_unlock(&hugetlb_lock);
351         for (i = 0; i < needed; i++) {
352                 page = alloc_buddy_huge_page(NULL, 0);
353                 if (!page) {
354                         /*
355                          * We were not able to allocate enough pages to
356                          * satisfy the entire reservation so we free what
357                          * we've allocated so far.
358                          */
359                         spin_lock(&hugetlb_lock);
360                         needed = 0;
361                         goto free;
362                 }
363
364                 list_add(&page->lru, &surplus_list);
365         }
366         allocated += needed;
367
368         /*
369          * After retaking hugetlb_lock, we need to recalculate 'needed'
370          * because either resv_huge_pages or free_huge_pages may have changed.
371          */
372         spin_lock(&hugetlb_lock);
373         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
374         if (needed > 0)
375                 goto retry;
376
377         /*
378          * The surplus_list now contains _at_least_ the number of extra pages
379          * needed to accomodate the reservation.  Add the appropriate number
380          * of pages to the hugetlb pool and free the extras back to the buddy
381          * allocator.  Commit the entire reservation here to prevent another
382          * process from stealing the pages as they are added to the pool but
383          * before they are reserved.
384          */
385         needed += allocated;
386         resv_huge_pages += delta;
387         ret = 0;
388 free:
389         /* Free the needed pages to the hugetlb pool */
390         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
391                 if ((--needed) < 0)
392                         break;
393                 list_del(&page->lru);
394                 enqueue_huge_page(page);
395         }
396
397         /* Free unnecessary surplus pages to the buddy allocator */
398         if (!list_empty(&surplus_list)) {
399                 spin_unlock(&hugetlb_lock);
400                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
401                         list_del(&page->lru);
402                         /*
403                          * The page has a reference count of zero already, so
404                          * call free_huge_page directly instead of using
405                          * put_page.  This must be done with hugetlb_lock
406                          * unlocked which is safe because free_huge_page takes
407                          * hugetlb_lock before deciding how to free the page.
408                          */
409                         free_huge_page(page);
410                 }
411                 spin_lock(&hugetlb_lock);
412         }
413
414         return ret;
415 }
416
417 /*
418  * When releasing a hugetlb pool reservation, any surplus pages that were
419  * allocated to satisfy the reservation must be explicitly freed if they were
420  * never used.
421  */
422 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
423 {
424         static int nid = -1;
425         struct page *page;
426         unsigned long nr_pages;
427
428         /*
429          * We want to release as many surplus pages as possible, spread
430          * evenly across all nodes. Iterate across all nodes until we
431          * can no longer free unreserved surplus pages. This occurs when
432          * the nodes with surplus pages have no free pages.
433          */
434         unsigned long remaining_iterations = num_online_nodes();
435
436         /* Uncommit the reservation */
437         resv_huge_pages -= unused_resv_pages;
438
439         nr_pages = min(unused_resv_pages, surplus_huge_pages);
440
441         while (remaining_iterations-- && nr_pages) {
442                 nid = next_node(nid, node_online_map);
443                 if (nid == MAX_NUMNODES)
444                         nid = first_node(node_online_map);
445
446                 if (!surplus_huge_pages_node[nid])
447                         continue;
448
449                 if (!list_empty(&hugepage_freelists[nid])) {
450                         page = list_entry(hugepage_freelists[nid].next,
451                                           struct page, lru);
452                         list_del(&page->lru);
453                         update_and_free_page(page);
454                         free_huge_pages--;
455                         free_huge_pages_node[nid]--;
456                         surplus_huge_pages--;
457                         surplus_huge_pages_node[nid]--;
458                         nr_pages--;
459                         remaining_iterations = num_online_nodes();
460                 }
461         }
462 }
463
464
465 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
466                                                 unsigned long addr)
467 {
468         struct page *page;
469
470         spin_lock(&hugetlb_lock);
471         page = dequeue_huge_page_vma(vma, addr);
472         spin_unlock(&hugetlb_lock);
473         return page ? page : ERR_PTR(-VM_FAULT_OOM);
474 }
475
476 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
477                                                 unsigned long addr)
478 {
479         struct page *page = NULL;
480
481         if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
482                 return ERR_PTR(-VM_FAULT_SIGBUS);
483
484         spin_lock(&hugetlb_lock);
485         if (free_huge_pages > resv_huge_pages)
486                 page = dequeue_huge_page_vma(vma, addr);
487         spin_unlock(&hugetlb_lock);
488         if (!page) {
489                 page = alloc_buddy_huge_page(vma, addr);
490                 if (!page) {
491                         hugetlb_put_quota(vma->vm_file->f_mapping, 1);
492                         return ERR_PTR(-VM_FAULT_OOM);
493                 }
494         }
495         return page;
496 }
497
498 static struct page *alloc_huge_page(struct vm_area_struct *vma,
499                                     unsigned long addr)
500 {
501         struct page *page;
502         struct address_space *mapping = vma->vm_file->f_mapping;
503
504         if (vma->vm_flags & VM_MAYSHARE)
505                 page = alloc_huge_page_shared(vma, addr);
506         else
507                 page = alloc_huge_page_private(vma, addr);
508
509         if (!IS_ERR(page)) {
510                 set_page_refcounted(page);
511                 set_page_private(page, (unsigned long) mapping);
512         }
513         return page;
514 }
515
516 static int __init hugetlb_init(void)
517 {
518         unsigned long i;
519
520         if (HPAGE_SHIFT == 0)
521                 return 0;
522
523         for (i = 0; i < MAX_NUMNODES; ++i)
524                 INIT_LIST_HEAD(&hugepage_freelists[i]);
525
526         hugetlb_next_nid = first_node(node_online_map);
527
528         for (i = 0; i < max_huge_pages; ++i) {
529                 if (!alloc_fresh_huge_page())
530                         break;
531         }
532         max_huge_pages = free_huge_pages = nr_huge_pages = i;
533         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
534         return 0;
535 }
536 module_init(hugetlb_init);
537
538 static int __init hugetlb_setup(char *s)
539 {
540         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
541                 max_huge_pages = 0;
542         return 1;
543 }
544 __setup("hugepages=", hugetlb_setup);
545
546 static unsigned int cpuset_mems_nr(unsigned int *array)
547 {
548         int node;
549         unsigned int nr = 0;
550
551         for_each_node_mask(node, cpuset_current_mems_allowed)
552                 nr += array[node];
553
554         return nr;
555 }
556
557 #ifdef CONFIG_SYSCTL
558 #ifdef CONFIG_HIGHMEM
559 static void try_to_free_low(unsigned long count)
560 {
561         int i;
562
563         for (i = 0; i < MAX_NUMNODES; ++i) {
564                 struct page *page, *next;
565                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
566                         if (count >= nr_huge_pages)
567                                 return;
568                         if (PageHighMem(page))
569                                 continue;
570                         list_del(&page->lru);
571                         update_and_free_page(page);
572                         free_huge_pages--;
573                         free_huge_pages_node[page_to_nid(page)]--;
574                 }
575         }
576 }
577 #else
578 static inline void try_to_free_low(unsigned long count)
579 {
580 }
581 #endif
582
583 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
584 static unsigned long set_max_huge_pages(unsigned long count)
585 {
586         unsigned long min_count, ret;
587
588         /*
589          * Increase the pool size
590          * First take pages out of surplus state.  Then make up the
591          * remaining difference by allocating fresh huge pages.
592          *
593          * We might race with alloc_buddy_huge_page() here and be unable
594          * to convert a surplus huge page to a normal huge page. That is
595          * not critical, though, it just means the overall size of the
596          * pool might be one hugepage larger than it needs to be, but
597          * within all the constraints specified by the sysctls.
598          */
599         spin_lock(&hugetlb_lock);
600         while (surplus_huge_pages && count > persistent_huge_pages) {
601                 if (!adjust_pool_surplus(-1))
602                         break;
603         }
604
605         while (count > persistent_huge_pages) {
606                 /*
607                  * If this allocation races such that we no longer need the
608                  * page, free_huge_page will handle it by freeing the page
609                  * and reducing the surplus.
610                  */
611                 spin_unlock(&hugetlb_lock);
612                 ret = alloc_fresh_huge_page();
613                 spin_lock(&hugetlb_lock);
614                 if (!ret)
615                         goto out;
616
617         }
618
619         /*
620          * Decrease the pool size
621          * First return free pages to the buddy allocator (being careful
622          * to keep enough around to satisfy reservations).  Then place
623          * pages into surplus state as needed so the pool will shrink
624          * to the desired size as pages become free.
625          *
626          * By placing pages into the surplus state independent of the
627          * overcommit value, we are allowing the surplus pool size to
628          * exceed overcommit. There are few sane options here. Since
629          * alloc_buddy_huge_page() is checking the global counter,
630          * though, we'll note that we're not allowed to exceed surplus
631          * and won't grow the pool anywhere else. Not until one of the
632          * sysctls are changed, or the surplus pages go out of use.
633          */
634         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
635         min_count = max(count, min_count);
636         try_to_free_low(min_count);
637         while (min_count < persistent_huge_pages) {
638                 struct page *page = dequeue_huge_page();
639                 if (!page)
640                         break;
641                 update_and_free_page(page);
642         }
643         while (count < persistent_huge_pages) {
644                 if (!adjust_pool_surplus(1))
645                         break;
646         }
647 out:
648         ret = persistent_huge_pages;
649         spin_unlock(&hugetlb_lock);
650         return ret;
651 }
652
653 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
654                            struct file *file, void __user *buffer,
655                            size_t *length, loff_t *ppos)
656 {
657         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
658         max_huge_pages = set_max_huge_pages(max_huge_pages);
659         return 0;
660 }
661
662 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
663                         struct file *file, void __user *buffer,
664                         size_t *length, loff_t *ppos)
665 {
666         proc_dointvec(table, write, file, buffer, length, ppos);
667         if (hugepages_treat_as_movable)
668                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
669         else
670                 htlb_alloc_mask = GFP_HIGHUSER;
671         return 0;
672 }
673
674 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
675                         struct file *file, void __user *buffer,
676                         size_t *length, loff_t *ppos)
677 {
678         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
679         spin_lock(&hugetlb_lock);
680         nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
681         spin_unlock(&hugetlb_lock);
682         return 0;
683 }
684
685 #endif /* CONFIG_SYSCTL */
686
687 int hugetlb_report_meminfo(char *buf)
688 {
689         return sprintf(buf,
690                         "HugePages_Total: %5lu\n"
691                         "HugePages_Free:  %5lu\n"
692                         "HugePages_Rsvd:  %5lu\n"
693                         "HugePages_Surp:  %5lu\n"
694                         "Hugepagesize:    %5lu kB\n",
695                         nr_huge_pages,
696                         free_huge_pages,
697                         resv_huge_pages,
698                         surplus_huge_pages,
699                         HPAGE_SIZE/1024);
700 }
701
702 int hugetlb_report_node_meminfo(int nid, char *buf)
703 {
704         return sprintf(buf,
705                 "Node %d HugePages_Total: %5u\n"
706                 "Node %d HugePages_Free:  %5u\n"
707                 "Node %d HugePages_Surp:  %5u\n",
708                 nid, nr_huge_pages_node[nid],
709                 nid, free_huge_pages_node[nid],
710                 nid, surplus_huge_pages_node[nid]);
711 }
712
713 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
714 unsigned long hugetlb_total_pages(void)
715 {
716         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
717 }
718
719 /*
720  * We cannot handle pagefaults against hugetlb pages at all.  They cause
721  * handle_mm_fault() to try to instantiate regular-sized pages in the
722  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
723  * this far.
724  */
725 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
726 {
727         BUG();
728         return 0;
729 }
730
731 struct vm_operations_struct hugetlb_vm_ops = {
732         .fault = hugetlb_vm_op_fault,
733 };
734
735 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
736                                 int writable)
737 {
738         pte_t entry;
739
740         if (writable) {
741                 entry =
742                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
743         } else {
744                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
745         }
746         entry = pte_mkyoung(entry);
747         entry = pte_mkhuge(entry);
748
749         return entry;
750 }
751
752 static void set_huge_ptep_writable(struct vm_area_struct *vma,
753                                    unsigned long address, pte_t *ptep)
754 {
755         pte_t entry;
756
757         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
758         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
759                 update_mmu_cache(vma, address, entry);
760         }
761 }
762
763
764 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
765                             struct vm_area_struct *vma)
766 {
767         pte_t *src_pte, *dst_pte, entry;
768         struct page *ptepage;
769         unsigned long addr;
770         int cow;
771
772         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
773
774         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
775                 src_pte = huge_pte_offset(src, addr);
776                 if (!src_pte)
777                         continue;
778                 dst_pte = huge_pte_alloc(dst, addr);
779                 if (!dst_pte)
780                         goto nomem;
781
782                 /* If the pagetables are shared don't copy or take references */
783                 if (dst_pte == src_pte)
784                         continue;
785
786                 spin_lock(&dst->page_table_lock);
787                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
788                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
789                         if (cow)
790                                 huge_ptep_set_wrprotect(src, addr, src_pte);
791                         entry = huge_ptep_get(src_pte);
792                         ptepage = pte_page(entry);
793                         get_page(ptepage);
794                         set_huge_pte_at(dst, addr, dst_pte, entry);
795                 }
796                 spin_unlock(&src->page_table_lock);
797                 spin_unlock(&dst->page_table_lock);
798         }
799         return 0;
800
801 nomem:
802         return -ENOMEM;
803 }
804
805 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
806                             unsigned long end)
807 {
808         struct mm_struct *mm = vma->vm_mm;
809         unsigned long address;
810         pte_t *ptep;
811         pte_t pte;
812         struct page *page;
813         struct page *tmp;
814         /*
815          * A page gathering list, protected by per file i_mmap_lock. The
816          * lock is used to avoid list corruption from multiple unmapping
817          * of the same page since we are using page->lru.
818          */
819         LIST_HEAD(page_list);
820
821         WARN_ON(!is_vm_hugetlb_page(vma));
822         BUG_ON(start & ~HPAGE_MASK);
823         BUG_ON(end & ~HPAGE_MASK);
824
825         spin_lock(&mm->page_table_lock);
826         for (address = start; address < end; address += HPAGE_SIZE) {
827                 ptep = huge_pte_offset(mm, address);
828                 if (!ptep)
829                         continue;
830
831                 if (huge_pmd_unshare(mm, &address, ptep))
832                         continue;
833
834                 pte = huge_ptep_get_and_clear(mm, address, ptep);
835                 if (huge_pte_none(pte))
836                         continue;
837
838                 page = pte_page(pte);
839                 if (pte_dirty(pte))
840                         set_page_dirty(page);
841                 list_add(&page->lru, &page_list);
842         }
843         spin_unlock(&mm->page_table_lock);
844         flush_tlb_range(vma, start, end);
845         list_for_each_entry_safe(page, tmp, &page_list, lru) {
846                 list_del(&page->lru);
847                 put_page(page);
848         }
849 }
850
851 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
852                           unsigned long end)
853 {
854         /*
855          * It is undesirable to test vma->vm_file as it should be non-null
856          * for valid hugetlb area. However, vm_file will be NULL in the error
857          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
858          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
859          * to clean up. Since no pte has actually been setup, it is safe to
860          * do nothing in this case.
861          */
862         if (vma->vm_file) {
863                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
864                 __unmap_hugepage_range(vma, start, end);
865                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
866         }
867 }
868
869 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
870                         unsigned long address, pte_t *ptep, pte_t pte)
871 {
872         struct page *old_page, *new_page;
873         int avoidcopy;
874
875         old_page = pte_page(pte);
876
877         /* If no-one else is actually using this page, avoid the copy
878          * and just make the page writable */
879         avoidcopy = (page_count(old_page) == 1);
880         if (avoidcopy) {
881                 set_huge_ptep_writable(vma, address, ptep);
882                 return 0;
883         }
884
885         page_cache_get(old_page);
886         new_page = alloc_huge_page(vma, address);
887
888         if (IS_ERR(new_page)) {
889                 page_cache_release(old_page);
890                 return -PTR_ERR(new_page);
891         }
892
893         spin_unlock(&mm->page_table_lock);
894         copy_huge_page(new_page, old_page, address, vma);
895         __SetPageUptodate(new_page);
896         spin_lock(&mm->page_table_lock);
897
898         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
899         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
900                 /* Break COW */
901                 huge_ptep_clear_flush(vma, address, ptep);
902                 set_huge_pte_at(mm, address, ptep,
903                                 make_huge_pte(vma, new_page, 1));
904                 /* Make the old page be freed below */
905                 new_page = old_page;
906         }
907         page_cache_release(new_page);
908         page_cache_release(old_page);
909         return 0;
910 }
911
912 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
913                         unsigned long address, pte_t *ptep, int write_access)
914 {
915         int ret = VM_FAULT_SIGBUS;
916         unsigned long idx;
917         unsigned long size;
918         struct page *page;
919         struct address_space *mapping;
920         pte_t new_pte;
921
922         mapping = vma->vm_file->f_mapping;
923         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
924                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
925
926         /*
927          * Use page lock to guard against racing truncation
928          * before we get page_table_lock.
929          */
930 retry:
931         page = find_lock_page(mapping, idx);
932         if (!page) {
933                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
934                 if (idx >= size)
935                         goto out;
936                 page = alloc_huge_page(vma, address);
937                 if (IS_ERR(page)) {
938                         ret = -PTR_ERR(page);
939                         goto out;
940                 }
941                 clear_huge_page(page, address);
942                 __SetPageUptodate(page);
943
944                 if (vma->vm_flags & VM_SHARED) {
945                         int err;
946                         struct inode *inode = mapping->host;
947
948                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
949                         if (err) {
950                                 put_page(page);
951                                 if (err == -EEXIST)
952                                         goto retry;
953                                 goto out;
954                         }
955
956                         spin_lock(&inode->i_lock);
957                         inode->i_blocks += BLOCKS_PER_HUGEPAGE;
958                         spin_unlock(&inode->i_lock);
959                 } else
960                         lock_page(page);
961         }
962
963         spin_lock(&mm->page_table_lock);
964         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
965         if (idx >= size)
966                 goto backout;
967
968         ret = 0;
969         if (!huge_pte_none(huge_ptep_get(ptep)))
970                 goto backout;
971
972         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
973                                 && (vma->vm_flags & VM_SHARED)));
974         set_huge_pte_at(mm, address, ptep, new_pte);
975
976         if (write_access && !(vma->vm_flags & VM_SHARED)) {
977                 /* Optimization, do the COW without a second fault */
978                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
979         }
980
981         spin_unlock(&mm->page_table_lock);
982         unlock_page(page);
983 out:
984         return ret;
985
986 backout:
987         spin_unlock(&mm->page_table_lock);
988         unlock_page(page);
989         put_page(page);
990         goto out;
991 }
992
993 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
994                         unsigned long address, int write_access)
995 {
996         pte_t *ptep;
997         pte_t entry;
998         int ret;
999         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
1000
1001         ptep = huge_pte_alloc(mm, address);
1002         if (!ptep)
1003                 return VM_FAULT_OOM;
1004
1005         /*
1006          * Serialize hugepage allocation and instantiation, so that we don't
1007          * get spurious allocation failures if two CPUs race to instantiate
1008          * the same page in the page cache.
1009          */
1010         mutex_lock(&hugetlb_instantiation_mutex);
1011         entry = huge_ptep_get(ptep);
1012         if (huge_pte_none(entry)) {
1013                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1014                 mutex_unlock(&hugetlb_instantiation_mutex);
1015                 return ret;
1016         }
1017
1018         ret = 0;
1019
1020         spin_lock(&mm->page_table_lock);
1021         /* Check for a racing update before calling hugetlb_cow */
1022         if (likely(pte_same(entry, huge_ptep_get(ptep))))
1023                 if (write_access && !pte_write(entry))
1024                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
1025         spin_unlock(&mm->page_table_lock);
1026         mutex_unlock(&hugetlb_instantiation_mutex);
1027
1028         return ret;
1029 }
1030
1031 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
1032                         struct page **pages, struct vm_area_struct **vmas,
1033                         unsigned long *position, int *length, int i,
1034                         int write)
1035 {
1036         unsigned long pfn_offset;
1037         unsigned long vaddr = *position;
1038         int remainder = *length;
1039
1040         spin_lock(&mm->page_table_lock);
1041         while (vaddr < vma->vm_end && remainder) {
1042                 pte_t *pte;
1043                 struct page *page;
1044
1045                 /*
1046                  * Some archs (sparc64, sh*) have multiple pte_ts to
1047                  * each hugepage.  We have to make * sure we get the
1048                  * first, for the page indexing below to work.
1049                  */
1050                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1051
1052                 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
1053                     (write && !pte_write(huge_ptep_get(pte)))) {
1054                         int ret;
1055
1056                         spin_unlock(&mm->page_table_lock);
1057                         ret = hugetlb_fault(mm, vma, vaddr, write);
1058                         spin_lock(&mm->page_table_lock);
1059                         if (!(ret & VM_FAULT_ERROR))
1060                                 continue;
1061
1062                         remainder = 0;
1063                         if (!i)
1064                                 i = -EFAULT;
1065                         break;
1066                 }
1067
1068                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1069                 page = pte_page(huge_ptep_get(pte));
1070 same_page:
1071                 if (pages) {
1072                         get_page(page);
1073                         pages[i] = page + pfn_offset;
1074                 }
1075
1076                 if (vmas)
1077                         vmas[i] = vma;
1078
1079                 vaddr += PAGE_SIZE;
1080                 ++pfn_offset;
1081                 --remainder;
1082                 ++i;
1083                 if (vaddr < vma->vm_end && remainder &&
1084                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1085                         /*
1086                          * We use pfn_offset to avoid touching the pageframes
1087                          * of this compound page.
1088                          */
1089                         goto same_page;
1090                 }
1091         }
1092         spin_unlock(&mm->page_table_lock);
1093         *length = remainder;
1094         *position = vaddr;
1095
1096         return i;
1097 }
1098
1099 void hugetlb_change_protection(struct vm_area_struct *vma,
1100                 unsigned long address, unsigned long end, pgprot_t newprot)
1101 {
1102         struct mm_struct *mm = vma->vm_mm;
1103         unsigned long start = address;
1104         pte_t *ptep;
1105         pte_t pte;
1106
1107         BUG_ON(address >= end);
1108         flush_cache_range(vma, address, end);
1109
1110         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1111         spin_lock(&mm->page_table_lock);
1112         for (; address < end; address += HPAGE_SIZE) {
1113                 ptep = huge_pte_offset(mm, address);
1114                 if (!ptep)
1115                         continue;
1116                 if (huge_pmd_unshare(mm, &address, ptep))
1117                         continue;
1118                 if (!huge_pte_none(huge_ptep_get(ptep))) {
1119                         pte = huge_ptep_get_and_clear(mm, address, ptep);
1120                         pte = pte_mkhuge(pte_modify(pte, newprot));
1121                         set_huge_pte_at(mm, address, ptep, pte);
1122                 }
1123         }
1124         spin_unlock(&mm->page_table_lock);
1125         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1126
1127         flush_tlb_range(vma, start, end);
1128 }
1129
1130 struct file_region {
1131         struct list_head link;
1132         long from;
1133         long to;
1134 };
1135
1136 static long region_add(struct list_head *head, long f, long t)
1137 {
1138         struct file_region *rg, *nrg, *trg;
1139
1140         /* Locate the region we are either in or before. */
1141         list_for_each_entry(rg, head, link)
1142                 if (f <= rg->to)
1143                         break;
1144
1145         /* Round our left edge to the current segment if it encloses us. */
1146         if (f > rg->from)
1147                 f = rg->from;
1148
1149         /* Check for and consume any regions we now overlap with. */
1150         nrg = rg;
1151         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1152                 if (&rg->link == head)
1153                         break;
1154                 if (rg->from > t)
1155                         break;
1156
1157                 /* If this area reaches higher then extend our area to
1158                  * include it completely.  If this is not the first area
1159                  * which we intend to reuse, free it. */
1160                 if (rg->to > t)
1161                         t = rg->to;
1162                 if (rg != nrg) {
1163                         list_del(&rg->link);
1164                         kfree(rg);
1165                 }
1166         }
1167         nrg->from = f;
1168         nrg->to = t;
1169         return 0;
1170 }
1171
1172 static long region_chg(struct list_head *head, long f, long t)
1173 {
1174         struct file_region *rg, *nrg;
1175         long chg = 0;
1176
1177         /* Locate the region we are before or in. */
1178         list_for_each_entry(rg, head, link)
1179                 if (f <= rg->to)
1180                         break;
1181
1182         /* If we are below the current region then a new region is required.
1183          * Subtle, allocate a new region at the position but make it zero
1184          * size such that we can guarantee to record the reservation. */
1185         if (&rg->link == head || t < rg->from) {
1186                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1187                 if (!nrg)
1188                         return -ENOMEM;
1189                 nrg->from = f;
1190                 nrg->to   = f;
1191                 INIT_LIST_HEAD(&nrg->link);
1192                 list_add(&nrg->link, rg->link.prev);
1193
1194                 return t - f;
1195         }
1196
1197         /* Round our left edge to the current segment if it encloses us. */
1198         if (f > rg->from)
1199                 f = rg->from;
1200         chg = t - f;
1201
1202         /* Check for and consume any regions we now overlap with. */
1203         list_for_each_entry(rg, rg->link.prev, link) {
1204                 if (&rg->link == head)
1205                         break;
1206                 if (rg->from > t)
1207                         return chg;
1208
1209                 /* We overlap with this area, if it extends futher than
1210                  * us then we must extend ourselves.  Account for its
1211                  * existing reservation. */
1212                 if (rg->to > t) {
1213                         chg += rg->to - t;
1214                         t = rg->to;
1215                 }
1216                 chg -= rg->to - rg->from;
1217         }
1218         return chg;
1219 }
1220
1221 static long region_truncate(struct list_head *head, long end)
1222 {
1223         struct file_region *rg, *trg;
1224         long chg = 0;
1225
1226         /* Locate the region we are either in or before. */
1227         list_for_each_entry(rg, head, link)
1228                 if (end <= rg->to)
1229                         break;
1230         if (&rg->link == head)
1231                 return 0;
1232
1233         /* If we are in the middle of a region then adjust it. */
1234         if (end > rg->from) {
1235                 chg = rg->to - end;
1236                 rg->to = end;
1237                 rg = list_entry(rg->link.next, typeof(*rg), link);
1238         }
1239
1240         /* Drop any remaining regions. */
1241         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1242                 if (&rg->link == head)
1243                         break;
1244                 chg += rg->to - rg->from;
1245                 list_del(&rg->link);
1246                 kfree(rg);
1247         }
1248         return chg;
1249 }
1250
1251 static int hugetlb_acct_memory(long delta)
1252 {
1253         int ret = -ENOMEM;
1254
1255         spin_lock(&hugetlb_lock);
1256         /*
1257          * When cpuset is configured, it breaks the strict hugetlb page
1258          * reservation as the accounting is done on a global variable. Such
1259          * reservation is completely rubbish in the presence of cpuset because
1260          * the reservation is not checked against page availability for the
1261          * current cpuset. Application can still potentially OOM'ed by kernel
1262          * with lack of free htlb page in cpuset that the task is in.
1263          * Attempt to enforce strict accounting with cpuset is almost
1264          * impossible (or too ugly) because cpuset is too fluid that
1265          * task or memory node can be dynamically moved between cpusets.
1266          *
1267          * The change of semantics for shared hugetlb mapping with cpuset is
1268          * undesirable. However, in order to preserve some of the semantics,
1269          * we fall back to check against current free page availability as
1270          * a best attempt and hopefully to minimize the impact of changing
1271          * semantics that cpuset has.
1272          */
1273         if (delta > 0) {
1274                 if (gather_surplus_pages(delta) < 0)
1275                         goto out;
1276
1277                 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1278                         return_unused_surplus_pages(delta);
1279                         goto out;
1280                 }
1281         }
1282
1283         ret = 0;
1284         if (delta < 0)
1285                 return_unused_surplus_pages((unsigned long) -delta);
1286
1287 out:
1288         spin_unlock(&hugetlb_lock);
1289         return ret;
1290 }
1291
1292 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1293 {
1294         long ret, chg;
1295
1296         chg = region_chg(&inode->i_mapping->private_list, from, to);
1297         if (chg < 0)
1298                 return chg;
1299
1300         if (hugetlb_get_quota(inode->i_mapping, chg))
1301                 return -ENOSPC;
1302         ret = hugetlb_acct_memory(chg);
1303         if (ret < 0) {
1304                 hugetlb_put_quota(inode->i_mapping, chg);
1305                 return ret;
1306         }
1307         region_add(&inode->i_mapping->private_list, from, to);
1308         return 0;
1309 }
1310
1311 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1312 {
1313         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1314
1315         spin_lock(&inode->i_lock);
1316         inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1317         spin_unlock(&inode->i_lock);
1318
1319         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1320         hugetlb_acct_memory(-(chg - freed));
1321 }