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