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