tcp: enforce tcp_min_snd_mss in tcp_mtu_probing()
[pandora-kernel.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/export.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31 #include <linux/audit.h>
32 #include <linux/khugepaged.h>
33
34 #include <asm/uaccess.h>
35 #include <asm/cacheflush.h>
36 #include <asm/tlb.h>
37 #include <asm/mmu_context.h>
38
39 #include "internal.h"
40
41 #ifndef arch_mmap_check
42 #define arch_mmap_check(addr, len, flags)       (0)
43 #endif
44
45 #ifndef arch_rebalance_pgtables
46 #define arch_rebalance_pgtables(addr, len)              (addr)
47 #endif
48
49 static void unmap_region(struct mm_struct *mm,
50                 struct vm_area_struct *vma, struct vm_area_struct *prev,
51                 unsigned long start, unsigned long end);
52
53 /*
54  * WARNING: the debugging will use recursive algorithms so never enable this
55  * unless you know what you are doing.
56  */
57 #undef DEBUG_MM_RB
58
59 /* description of effects of mapping type and prot in current implementation.
60  * this is due to the limited x86 page protection hardware.  The expected
61  * behavior is in parens:
62  *
63  * map_type     prot
64  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
65  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
66  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
67  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
68  *              
69  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
70  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
71  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
72  *
73  */
74 pgprot_t protection_map[16] = {
75         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
76         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
77 };
78
79 pgprot_t vm_get_page_prot(unsigned long vm_flags)
80 {
81         return __pgprot(pgprot_val(protection_map[vm_flags &
82                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
83                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
84 }
85 EXPORT_SYMBOL(vm_get_page_prot);
86
87 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
88 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
89 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
90 /*
91  * Make sure vm_committed_as in one cacheline and not cacheline shared with
92  * other variables. It can be updated by several CPUs frequently.
93  */
94 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
95
96 /*
97  * Check that a process has enough memory to allocate a new virtual
98  * mapping. 0 means there is enough memory for the allocation to
99  * succeed and -ENOMEM implies there is not.
100  *
101  * We currently support three overcommit policies, which are set via the
102  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
103  *
104  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
105  * Additional code 2002 Jul 20 by Robert Love.
106  *
107  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
108  *
109  * Note this is a helper function intended to be used by LSMs which
110  * wish to use this logic.
111  */
112 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
113 {
114         long free, allowed;
115
116         vm_acct_memory(pages);
117
118         /*
119          * Sometimes we want to use more memory than we have
120          */
121         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
122                 return 0;
123
124         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
125                 free = global_page_state(NR_FREE_PAGES);
126                 free += global_page_state(NR_FILE_PAGES);
127
128                 /*
129                  * shmem pages shouldn't be counted as free in this
130                  * case, they can't be purged, only swapped out, and
131                  * that won't affect the overall amount of available
132                  * memory in the system.
133                  */
134                 free -= global_page_state(NR_SHMEM);
135
136                 free += nr_swap_pages;
137
138                 /*
139                  * Any slabs which are created with the
140                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
141                  * which are reclaimable, under pressure.  The dentry
142                  * cache and most inode caches should fall into this
143                  */
144                 free += global_page_state(NR_SLAB_RECLAIMABLE);
145
146                 /*
147                  * Leave reserved pages. The pages are not for anonymous pages.
148                  */
149                 if (free <= totalreserve_pages)
150                         goto error;
151                 else
152                         free -= totalreserve_pages;
153
154                 /*
155                  * Leave the last 3% for root
156                  */
157                 if (!cap_sys_admin)
158                         free -= free / 32;
159
160                 if (free > pages)
161                         return 0;
162
163                 goto error;
164         }
165
166         allowed = (totalram_pages - hugetlb_total_pages())
167                 * sysctl_overcommit_ratio / 100;
168         /*
169          * Leave the last 3% for root
170          */
171         if (!cap_sys_admin)
172                 allowed -= allowed / 32;
173         allowed += total_swap_pages;
174
175         /* Don't let a single process grow too big:
176            leave 3% of the size of this process for other processes */
177         if (mm)
178                 allowed -= mm->total_vm / 32;
179
180         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
181                 return 0;
182 error:
183         vm_unacct_memory(pages);
184
185         return -ENOMEM;
186 }
187
188 /*
189  * Requires inode->i_mapping->i_mmap_mutex
190  */
191 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
192                 struct file *file, struct address_space *mapping)
193 {
194         if (vma->vm_flags & VM_DENYWRITE)
195                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
196         if (vma->vm_flags & VM_SHARED)
197                 mapping_unmap_writable(mapping);
198
199         flush_dcache_mmap_lock(mapping);
200         if (unlikely(vma->vm_flags & VM_NONLINEAR))
201                 list_del_init(&vma->shared.vm_set.list);
202         else
203                 vma_prio_tree_remove(vma, &mapping->i_mmap);
204         flush_dcache_mmap_unlock(mapping);
205 }
206
207 /*
208  * Unlink a file-based vm structure from its prio_tree, to hide
209  * vma from rmap and vmtruncate before freeing its page tables.
210  */
211 void unlink_file_vma(struct vm_area_struct *vma)
212 {
213         struct file *file = vma->vm_file;
214
215         if (file) {
216                 struct address_space *mapping = file->f_mapping;
217                 mutex_lock(&mapping->i_mmap_mutex);
218                 __remove_shared_vm_struct(vma, file, mapping);
219                 mutex_unlock(&mapping->i_mmap_mutex);
220         }
221 }
222
223 /*
224  * Close a vm structure and free it, returning the next.
225  */
226 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
227 {
228         struct vm_area_struct *next = vma->vm_next;
229
230         might_sleep();
231         if (vma->vm_ops && vma->vm_ops->close)
232                 vma->vm_ops->close(vma);
233         if (vma->vm_file) {
234                 fput(vma->vm_file);
235                 if (vma->vm_flags & VM_EXECUTABLE)
236                         removed_exe_file_vma(vma->vm_mm);
237         }
238         mpol_put(vma_policy(vma));
239         kmem_cache_free(vm_area_cachep, vma);
240         return next;
241 }
242
243 SYSCALL_DEFINE1(brk, unsigned long, brk)
244 {
245         unsigned long rlim, retval;
246         unsigned long newbrk, oldbrk;
247         struct mm_struct *mm = current->mm;
248         struct vm_area_struct *next;
249         unsigned long min_brk;
250
251         down_write(&mm->mmap_sem);
252
253 #ifdef CONFIG_COMPAT_BRK
254         /*
255          * CONFIG_COMPAT_BRK can still be overridden by setting
256          * randomize_va_space to 2, which will still cause mm->start_brk
257          * to be arbitrarily shifted
258          */
259         if (current->brk_randomized)
260                 min_brk = mm->start_brk;
261         else
262                 min_brk = mm->end_data;
263 #else
264         min_brk = mm->start_brk;
265 #endif
266         if (brk < min_brk)
267                 goto out;
268
269         /*
270          * Check against rlimit here. If this check is done later after the test
271          * of oldbrk with newbrk then it can escape the test and let the data
272          * segment grow beyond its set limit the in case where the limit is
273          * not page aligned -Ram Gupta
274          */
275         rlim = rlimit(RLIMIT_DATA);
276         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
277                         (mm->end_data - mm->start_data) > rlim)
278                 goto out;
279
280         newbrk = PAGE_ALIGN(brk);
281         oldbrk = PAGE_ALIGN(mm->brk);
282         if (oldbrk == newbrk)
283                 goto set_brk;
284
285         /* Always allow shrinking brk. */
286         if (brk <= mm->brk) {
287                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
288                         goto set_brk;
289                 goto out;
290         }
291
292         /* Check against existing mmap mappings. */
293         next = find_vma(mm, oldbrk);
294         if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
295                 goto out;
296
297         /* Ok, looks good - let it rip. */
298         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
299                 goto out;
300 set_brk:
301         mm->brk = brk;
302 out:
303         retval = mm->brk;
304         up_write(&mm->mmap_sem);
305         return retval;
306 }
307
308 #ifdef DEBUG_MM_RB
309 static int browse_rb(struct rb_root *root)
310 {
311         int i = 0, j;
312         struct rb_node *nd, *pn = NULL;
313         unsigned long prev = 0, pend = 0;
314
315         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
316                 struct vm_area_struct *vma;
317                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
318                 if (vma->vm_start < prev)
319                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
320                 if (vma->vm_start < pend)
321                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
322                 if (vma->vm_start > vma->vm_end)
323                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
324                 i++;
325                 pn = nd;
326                 prev = vma->vm_start;
327                 pend = vma->vm_end;
328         }
329         j = 0;
330         for (nd = pn; nd; nd = rb_prev(nd)) {
331                 j++;
332         }
333         if (i != j)
334                 printk("backwards %d, forwards %d\n", j, i), i = 0;
335         return i;
336 }
337
338 void validate_mm(struct mm_struct *mm)
339 {
340         int bug = 0;
341         int i = 0;
342         struct vm_area_struct *tmp = mm->mmap;
343         while (tmp) {
344                 tmp = tmp->vm_next;
345                 i++;
346         }
347         if (i != mm->map_count)
348                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
349         i = browse_rb(&mm->mm_rb);
350         if (i != mm->map_count)
351                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
352         BUG_ON(bug);
353 }
354 #else
355 #define validate_mm(mm) do { } while (0)
356 #endif
357
358 static struct vm_area_struct *
359 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
360                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
361                 struct rb_node ** rb_parent)
362 {
363         struct vm_area_struct * vma;
364         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
365
366         __rb_link = &mm->mm_rb.rb_node;
367         rb_prev = __rb_parent = NULL;
368         vma = NULL;
369
370         while (*__rb_link) {
371                 struct vm_area_struct *vma_tmp;
372
373                 __rb_parent = *__rb_link;
374                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
375
376                 if (vma_tmp->vm_end > addr) {
377                         vma = vma_tmp;
378                         if (vma_tmp->vm_start <= addr)
379                                 break;
380                         __rb_link = &__rb_parent->rb_left;
381                 } else {
382                         rb_prev = __rb_parent;
383                         __rb_link = &__rb_parent->rb_right;
384                 }
385         }
386
387         *pprev = NULL;
388         if (rb_prev)
389                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
390         *rb_link = __rb_link;
391         *rb_parent = __rb_parent;
392         return vma;
393 }
394
395 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
396                 struct rb_node **rb_link, struct rb_node *rb_parent)
397 {
398         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
399         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
400 }
401
402 static void __vma_link_file(struct vm_area_struct *vma)
403 {
404         struct file *file;
405
406         file = vma->vm_file;
407         if (file) {
408                 struct address_space *mapping = file->f_mapping;
409
410                 if (vma->vm_flags & VM_DENYWRITE)
411                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
412                 if (vma->vm_flags & VM_SHARED)
413                         atomic_inc(&mapping->i_mmap_writable);
414
415                 flush_dcache_mmap_lock(mapping);
416                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
417                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
418                 else
419                         vma_prio_tree_insert(vma, &mapping->i_mmap);
420                 flush_dcache_mmap_unlock(mapping);
421         }
422 }
423
424 static void
425 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
426         struct vm_area_struct *prev, struct rb_node **rb_link,
427         struct rb_node *rb_parent)
428 {
429         __vma_link_list(mm, vma, prev, rb_parent);
430         __vma_link_rb(mm, vma, rb_link, rb_parent);
431 }
432
433 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
434                         struct vm_area_struct *prev, struct rb_node **rb_link,
435                         struct rb_node *rb_parent)
436 {
437         struct address_space *mapping = NULL;
438
439         if (vma->vm_file)
440                 mapping = vma->vm_file->f_mapping;
441
442         if (mapping)
443                 mutex_lock(&mapping->i_mmap_mutex);
444
445         __vma_link(mm, vma, prev, rb_link, rb_parent);
446         __vma_link_file(vma);
447
448         if (mapping)
449                 mutex_unlock(&mapping->i_mmap_mutex);
450
451         mm->map_count++;
452         validate_mm(mm);
453 }
454
455 /*
456  * Helper for vma_adjust in the split_vma insert case:
457  * insert vm structure into list and rbtree and anon_vma,
458  * but it has already been inserted into prio_tree earlier.
459  */
460 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
461 {
462         struct vm_area_struct *__vma, *prev;
463         struct rb_node **rb_link, *rb_parent;
464
465         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
466         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
467         __vma_link(mm, vma, prev, rb_link, rb_parent);
468         mm->map_count++;
469 }
470
471 static inline void
472 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
473                 struct vm_area_struct *prev)
474 {
475         struct vm_area_struct *next = vma->vm_next;
476
477         prev->vm_next = next;
478         if (next)
479                 next->vm_prev = prev;
480         rb_erase(&vma->vm_rb, &mm->mm_rb);
481         if (mm->mmap_cache == vma)
482                 mm->mmap_cache = prev;
483 }
484
485 /*
486  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
487  * is already present in an i_mmap tree without adjusting the tree.
488  * The following helper function should be used when such adjustments
489  * are necessary.  The "insert" vma (if any) is to be inserted
490  * before we drop the necessary locks.
491  */
492 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
493         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
494 {
495         struct mm_struct *mm = vma->vm_mm;
496         struct vm_area_struct *next = vma->vm_next;
497         struct vm_area_struct *importer = NULL;
498         struct address_space *mapping = NULL;
499         struct prio_tree_root *root = NULL;
500         struct anon_vma *anon_vma = NULL;
501         struct file *file = vma->vm_file;
502         long adjust_next = 0;
503         int remove_next = 0;
504
505         if (next && !insert) {
506                 struct vm_area_struct *exporter = NULL;
507
508                 if (end >= next->vm_end) {
509                         /*
510                          * vma expands, overlapping all the next, and
511                          * perhaps the one after too (mprotect case 6).
512                          */
513 again:                  remove_next = 1 + (end > next->vm_end);
514                         end = next->vm_end;
515                         exporter = next;
516                         importer = vma;
517                 } else if (end > next->vm_start) {
518                         /*
519                          * vma expands, overlapping part of the next:
520                          * mprotect case 5 shifting the boundary up.
521                          */
522                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
523                         exporter = next;
524                         importer = vma;
525                 } else if (end < vma->vm_end) {
526                         /*
527                          * vma shrinks, and !insert tells it's not
528                          * split_vma inserting another: so it must be
529                          * mprotect case 4 shifting the boundary down.
530                          */
531                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
532                         exporter = vma;
533                         importer = next;
534                 }
535
536                 /*
537                  * Easily overlooked: when mprotect shifts the boundary,
538                  * make sure the expanding vma has anon_vma set if the
539                  * shrinking vma had, to cover any anon pages imported.
540                  */
541                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
542                         int error;
543
544                         importer->anon_vma = exporter->anon_vma;
545                         error = anon_vma_clone(importer, exporter);
546                         if (error)
547                                 return error;
548                 }
549         }
550
551         if (file) {
552                 mapping = file->f_mapping;
553                 if (!(vma->vm_flags & VM_NONLINEAR))
554                         root = &mapping->i_mmap;
555                 mutex_lock(&mapping->i_mmap_mutex);
556                 if (insert) {
557                         /*
558                          * Put into prio_tree now, so instantiated pages
559                          * are visible to arm/parisc __flush_dcache_page
560                          * throughout; but we cannot insert into address
561                          * space until vma start or end is updated.
562                          */
563                         __vma_link_file(insert);
564                 }
565         }
566
567         vma_adjust_trans_huge(vma, start, end, adjust_next);
568
569         /*
570          * When changing only vma->vm_end, we don't really need anon_vma
571          * lock. This is a fairly rare case by itself, but the anon_vma
572          * lock may be shared between many sibling processes.  Skipping
573          * the lock for brk adjustments makes a difference sometimes.
574          */
575         if (vma->anon_vma && (importer || start != vma->vm_start)) {
576                 anon_vma = vma->anon_vma;
577                 anon_vma_lock(anon_vma);
578         }
579
580         if (root) {
581                 flush_dcache_mmap_lock(mapping);
582                 vma_prio_tree_remove(vma, root);
583                 if (adjust_next)
584                         vma_prio_tree_remove(next, root);
585         }
586
587         vma->vm_start = start;
588         vma->vm_end = end;
589         vma->vm_pgoff = pgoff;
590         if (adjust_next) {
591                 next->vm_start += adjust_next << PAGE_SHIFT;
592                 next->vm_pgoff += adjust_next;
593         }
594
595         if (root) {
596                 if (adjust_next)
597                         vma_prio_tree_insert(next, root);
598                 vma_prio_tree_insert(vma, root);
599                 flush_dcache_mmap_unlock(mapping);
600         }
601
602         if (remove_next) {
603                 /*
604                  * vma_merge has merged next into vma, and needs
605                  * us to remove next before dropping the locks.
606                  */
607                 __vma_unlink(mm, next, vma);
608                 if (file)
609                         __remove_shared_vm_struct(next, file, mapping);
610         } else if (insert) {
611                 /*
612                  * split_vma has split insert from vma, and needs
613                  * us to insert it before dropping the locks
614                  * (it may either follow vma or precede it).
615                  */
616                 __insert_vm_struct(mm, insert);
617         }
618
619         if (anon_vma)
620                 anon_vma_unlock(anon_vma);
621         if (mapping)
622                 mutex_unlock(&mapping->i_mmap_mutex);
623
624         if (remove_next) {
625                 if (file) {
626                         fput(file);
627                         if (next->vm_flags & VM_EXECUTABLE)
628                                 removed_exe_file_vma(mm);
629                 }
630                 if (next->anon_vma)
631                         anon_vma_merge(vma, next);
632                 mm->map_count--;
633                 mpol_put(vma_policy(next));
634                 kmem_cache_free(vm_area_cachep, next);
635                 /*
636                  * In mprotect's case 6 (see comments on vma_merge),
637                  * we must remove another next too. It would clutter
638                  * up the code too much to do both in one go.
639                  */
640                 if (remove_next == 2) {
641                         next = vma->vm_next;
642                         goto again;
643                 }
644         }
645
646         validate_mm(mm);
647
648         return 0;
649 }
650
651 /*
652  * If the vma has a ->close operation then the driver probably needs to release
653  * per-vma resources, so we don't attempt to merge those.
654  */
655 static inline int is_mergeable_vma(struct vm_area_struct *vma,
656                         struct file *file, unsigned long vm_flags)
657 {
658         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
659         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
660                 return 0;
661         if (vma->vm_file != file)
662                 return 0;
663         if (vma->vm_ops && vma->vm_ops->close)
664                 return 0;
665         return 1;
666 }
667
668 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
669                                         struct anon_vma *anon_vma2,
670                                         struct vm_area_struct *vma)
671 {
672         /*
673          * The list_is_singular() test is to avoid merging VMA cloned from
674          * parents. This can improve scalability caused by anon_vma lock.
675          */
676         if ((!anon_vma1 || !anon_vma2) && (!vma ||
677                 list_is_singular(&vma->anon_vma_chain)))
678                 return 1;
679         return anon_vma1 == anon_vma2;
680 }
681
682 /*
683  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
684  * in front of (at a lower virtual address and file offset than) the vma.
685  *
686  * We cannot merge two vmas if they have differently assigned (non-NULL)
687  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
688  *
689  * We don't check here for the merged mmap wrapping around the end of pagecache
690  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
691  * wrap, nor mmaps which cover the final page at index -1UL.
692  */
693 static int
694 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
695         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
696 {
697         if (is_mergeable_vma(vma, file, vm_flags) &&
698             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
699                 if (vma->vm_pgoff == vm_pgoff)
700                         return 1;
701         }
702         return 0;
703 }
704
705 /*
706  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
707  * beyond (at a higher virtual address and file offset than) the vma.
708  *
709  * We cannot merge two vmas if they have differently assigned (non-NULL)
710  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
711  */
712 static int
713 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
714         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
715 {
716         if (is_mergeable_vma(vma, file, vm_flags) &&
717             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
718                 pgoff_t vm_pglen;
719                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
720                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
721                         return 1;
722         }
723         return 0;
724 }
725
726 /*
727  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
728  * whether that can be merged with its predecessor or its successor.
729  * Or both (it neatly fills a hole).
730  *
731  * In most cases - when called for mmap, brk or mremap - [addr,end) is
732  * certain not to be mapped by the time vma_merge is called; but when
733  * called for mprotect, it is certain to be already mapped (either at
734  * an offset within prev, or at the start of next), and the flags of
735  * this area are about to be changed to vm_flags - and the no-change
736  * case has already been eliminated.
737  *
738  * The following mprotect cases have to be considered, where AAAA is
739  * the area passed down from mprotect_fixup, never extending beyond one
740  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
741  *
742  *     AAAA             AAAA                AAAA          AAAA
743  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
744  *    cannot merge    might become    might become    might become
745  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
746  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
747  *    mremap move:                                    PPPPNNNNNNNN 8
748  *        AAAA
749  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
750  *    might become    case 1 below    case 2 below    case 3 below
751  *
752  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
753  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
754  */
755 struct vm_area_struct *vma_merge(struct mm_struct *mm,
756                         struct vm_area_struct *prev, unsigned long addr,
757                         unsigned long end, unsigned long vm_flags,
758                         struct anon_vma *anon_vma, struct file *file,
759                         pgoff_t pgoff, struct mempolicy *policy)
760 {
761         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
762         struct vm_area_struct *area, *next;
763         int err;
764
765         /*
766          * We later require that vma->vm_flags == vm_flags,
767          * so this tests vma->vm_flags & VM_SPECIAL, too.
768          */
769         if (vm_flags & VM_SPECIAL)
770                 return NULL;
771
772         if (prev)
773                 next = prev->vm_next;
774         else
775                 next = mm->mmap;
776         area = next;
777         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
778                 next = next->vm_next;
779
780         /*
781          * Can it merge with the predecessor?
782          */
783         if (prev && prev->vm_end == addr &&
784                         mpol_equal(vma_policy(prev), policy) &&
785                         can_vma_merge_after(prev, vm_flags,
786                                                 anon_vma, file, pgoff)) {
787                 /*
788                  * OK, it can.  Can we now merge in the successor as well?
789                  */
790                 if (next && end == next->vm_start &&
791                                 mpol_equal(policy, vma_policy(next)) &&
792                                 can_vma_merge_before(next, vm_flags,
793                                         anon_vma, file, pgoff+pglen) &&
794                                 is_mergeable_anon_vma(prev->anon_vma,
795                                                       next->anon_vma, NULL)) {
796                                                         /* cases 1, 6 */
797                         err = vma_adjust(prev, prev->vm_start,
798                                 next->vm_end, prev->vm_pgoff, NULL);
799                 } else                                  /* cases 2, 5, 7 */
800                         err = vma_adjust(prev, prev->vm_start,
801                                 end, prev->vm_pgoff, NULL);
802                 if (err)
803                         return NULL;
804                 khugepaged_enter_vma_merge(prev, vm_flags);
805                 return prev;
806         }
807
808         /*
809          * Can this new request be merged in front of next?
810          */
811         if (next && end == next->vm_start &&
812                         mpol_equal(policy, vma_policy(next)) &&
813                         can_vma_merge_before(next, vm_flags,
814                                         anon_vma, file, pgoff+pglen)) {
815                 if (prev && addr < prev->vm_end)        /* case 4 */
816                         err = vma_adjust(prev, prev->vm_start,
817                                 addr, prev->vm_pgoff, NULL);
818                 else                                    /* cases 3, 8 */
819                         err = vma_adjust(area, addr, next->vm_end,
820                                 next->vm_pgoff - pglen, NULL);
821                 if (err)
822                         return NULL;
823                 khugepaged_enter_vma_merge(area, vm_flags);
824                 return area;
825         }
826
827         return NULL;
828 }
829
830 /*
831  * Rough compatbility check to quickly see if it's even worth looking
832  * at sharing an anon_vma.
833  *
834  * They need to have the same vm_file, and the flags can only differ
835  * in things that mprotect may change.
836  *
837  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
838  * we can merge the two vma's. For example, we refuse to merge a vma if
839  * there is a vm_ops->close() function, because that indicates that the
840  * driver is doing some kind of reference counting. But that doesn't
841  * really matter for the anon_vma sharing case.
842  */
843 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
844 {
845         return a->vm_end == b->vm_start &&
846                 mpol_equal(vma_policy(a), vma_policy(b)) &&
847                 a->vm_file == b->vm_file &&
848                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
849                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
850 }
851
852 /*
853  * Do some basic sanity checking to see if we can re-use the anon_vma
854  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
855  * the same as 'old', the other will be the new one that is trying
856  * to share the anon_vma.
857  *
858  * NOTE! This runs with mm_sem held for reading, so it is possible that
859  * the anon_vma of 'old' is concurrently in the process of being set up
860  * by another page fault trying to merge _that_. But that's ok: if it
861  * is being set up, that automatically means that it will be a singleton
862  * acceptable for merging, so we can do all of this optimistically. But
863  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
864  *
865  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
866  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
867  * is to return an anon_vma that is "complex" due to having gone through
868  * a fork).
869  *
870  * We also make sure that the two vma's are compatible (adjacent,
871  * and with the same memory policies). That's all stable, even with just
872  * a read lock on the mm_sem.
873  */
874 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
875 {
876         if (anon_vma_compatible(a, b)) {
877                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
878
879                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
880                         return anon_vma;
881         }
882         return NULL;
883 }
884
885 /*
886  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
887  * neighbouring vmas for a suitable anon_vma, before it goes off
888  * to allocate a new anon_vma.  It checks because a repetitive
889  * sequence of mprotects and faults may otherwise lead to distinct
890  * anon_vmas being allocated, preventing vma merge in subsequent
891  * mprotect.
892  */
893 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
894 {
895         struct anon_vma *anon_vma;
896         struct vm_area_struct *near;
897
898         near = vma->vm_next;
899         if (!near)
900                 goto try_prev;
901
902         anon_vma = reusable_anon_vma(near, vma, near);
903         if (anon_vma)
904                 return anon_vma;
905 try_prev:
906         near = vma->vm_prev;
907         if (!near)
908                 goto none;
909
910         anon_vma = reusable_anon_vma(near, near, vma);
911         if (anon_vma)
912                 return anon_vma;
913 none:
914         /*
915          * There's no absolute need to look only at touching neighbours:
916          * we could search further afield for "compatible" anon_vmas.
917          * But it would probably just be a waste of time searching,
918          * or lead to too many vmas hanging off the same anon_vma.
919          * We're trying to allow mprotect remerging later on,
920          * not trying to minimize memory used for anon_vmas.
921          */
922         return NULL;
923 }
924
925 #ifdef CONFIG_PROC_FS
926 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
927                                                 struct file *file, long pages)
928 {
929         const unsigned long stack_flags
930                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
931
932         if (file) {
933                 mm->shared_vm += pages;
934                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
935                         mm->exec_vm += pages;
936         } else if (flags & stack_flags)
937                 mm->stack_vm += pages;
938         if (flags & (VM_RESERVED|VM_IO))
939                 mm->reserved_vm += pages;
940 }
941 #endif /* CONFIG_PROC_FS */
942
943 /*
944  * The caller must hold down_write(&current->mm->mmap_sem).
945  */
946
947 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
948                         unsigned long len, unsigned long prot,
949                         unsigned long flags, unsigned long pgoff)
950 {
951         struct mm_struct * mm = current->mm;
952         struct inode *inode;
953         vm_flags_t vm_flags;
954         int error;
955         unsigned long reqprot = prot;
956
957         /*
958          * Does the application expect PROT_READ to imply PROT_EXEC?
959          *
960          * (the exception is when the underlying filesystem is noexec
961          *  mounted, in which case we dont add PROT_EXEC.)
962          */
963         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
964                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
965                         prot |= PROT_EXEC;
966
967         if (!len)
968                 return -EINVAL;
969
970         if (!(flags & MAP_FIXED))
971                 addr = round_hint_to_min(addr);
972
973         /* Careful about overflows.. */
974         len = PAGE_ALIGN(len);
975         if (!len)
976                 return -ENOMEM;
977
978         /* offset overflow? */
979         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
980                return -EOVERFLOW;
981
982         /* Too many mappings? */
983         if (mm->map_count > sysctl_max_map_count)
984                 return -ENOMEM;
985
986         /* Obtain the address to map to. we verify (or select) it and ensure
987          * that it represents a valid section of the address space.
988          */
989         addr = get_unmapped_area(file, addr, len, pgoff, flags);
990         if (addr & ~PAGE_MASK)
991                 return addr;
992
993         /* Do simple checking here so the lower-level routines won't have
994          * to. we assume access permissions have been handled by the open
995          * of the memory object, so we don't do any here.
996          */
997         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
998                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
999
1000         if (flags & MAP_LOCKED)
1001                 if (!can_do_mlock())
1002                         return -EPERM;
1003
1004         /* mlock MCL_FUTURE? */
1005         if (vm_flags & VM_LOCKED) {
1006                 unsigned long locked, lock_limit;
1007                 locked = len >> PAGE_SHIFT;
1008                 locked += mm->locked_vm;
1009                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1010                 lock_limit >>= PAGE_SHIFT;
1011                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1012                         return -EAGAIN;
1013         }
1014
1015         inode = file ? file->f_path.dentry->d_inode : NULL;
1016
1017         if (file) {
1018                 switch (flags & MAP_TYPE) {
1019                 case MAP_SHARED:
1020                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1021                                 return -EACCES;
1022
1023                         /*
1024                          * Make sure we don't allow writing to an append-only
1025                          * file..
1026                          */
1027                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1028                                 return -EACCES;
1029
1030                         /*
1031                          * Make sure there are no mandatory locks on the file.
1032                          */
1033                         if (locks_verify_locked(inode))
1034                                 return -EAGAIN;
1035
1036                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1037                         if (!(file->f_mode & FMODE_WRITE))
1038                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1039
1040                         /* fall through */
1041                 case MAP_PRIVATE:
1042                         if (!(file->f_mode & FMODE_READ))
1043                                 return -EACCES;
1044                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1045                                 if (vm_flags & VM_EXEC)
1046                                         return -EPERM;
1047                                 vm_flags &= ~VM_MAYEXEC;
1048                         }
1049
1050                         if (!file->f_op || !file->f_op->mmap)
1051                                 return -ENODEV;
1052                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1053                                 return -EINVAL;
1054                         break;
1055
1056                 default:
1057                         return -EINVAL;
1058                 }
1059         } else {
1060                 switch (flags & MAP_TYPE) {
1061                 case MAP_SHARED:
1062                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1063                                 return -EINVAL;
1064                         /*
1065                          * Ignore pgoff.
1066                          */
1067                         pgoff = 0;
1068                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1069                         break;
1070                 case MAP_PRIVATE:
1071                         /*
1072                          * Set pgoff according to addr for anon_vma.
1073                          */
1074                         pgoff = addr >> PAGE_SHIFT;
1075                         break;
1076                 default:
1077                         return -EINVAL;
1078                 }
1079         }
1080
1081         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1082         if (error)
1083                 return error;
1084
1085         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1086 }
1087 EXPORT_SYMBOL(do_mmap_pgoff);
1088
1089 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1090                 unsigned long, prot, unsigned long, flags,
1091                 unsigned long, fd, unsigned long, pgoff)
1092 {
1093         struct file *file = NULL;
1094         unsigned long retval = -EBADF;
1095
1096         if (!(flags & MAP_ANONYMOUS)) {
1097                 audit_mmap_fd(fd, flags);
1098                 if (unlikely(flags & MAP_HUGETLB))
1099                         return -EINVAL;
1100                 file = fget(fd);
1101                 if (!file)
1102                         goto out;
1103         } else if (flags & MAP_HUGETLB) {
1104                 struct user_struct *user = NULL;
1105                 /*
1106                  * VM_NORESERVE is used because the reservations will be
1107                  * taken when vm_ops->mmap() is called
1108                  * A dummy user value is used because we are not locking
1109                  * memory so no accounting is necessary
1110                  */
1111                 len = ALIGN(len, huge_page_size(&default_hstate));
1112                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1113                                                 &user, HUGETLB_ANONHUGE_INODE);
1114                 if (IS_ERR(file))
1115                         return PTR_ERR(file);
1116         }
1117
1118         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1119
1120         down_write(&current->mm->mmap_sem);
1121         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1122         up_write(&current->mm->mmap_sem);
1123
1124         if (file)
1125                 fput(file);
1126 out:
1127         return retval;
1128 }
1129
1130 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1131 struct mmap_arg_struct {
1132         unsigned long addr;
1133         unsigned long len;
1134         unsigned long prot;
1135         unsigned long flags;
1136         unsigned long fd;
1137         unsigned long offset;
1138 };
1139
1140 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1141 {
1142         struct mmap_arg_struct a;
1143
1144         if (copy_from_user(&a, arg, sizeof(a)))
1145                 return -EFAULT;
1146         if (a.offset & ~PAGE_MASK)
1147                 return -EINVAL;
1148
1149         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1150                               a.offset >> PAGE_SHIFT);
1151 }
1152 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1153
1154 /*
1155  * Some shared mappigns will want the pages marked read-only
1156  * to track write events. If so, we'll downgrade vm_page_prot
1157  * to the private version (using protection_map[] without the
1158  * VM_SHARED bit).
1159  */
1160 int vma_wants_writenotify(struct vm_area_struct *vma)
1161 {
1162         vm_flags_t vm_flags = vma->vm_flags;
1163
1164         /* If it was private or non-writable, the write bit is already clear */
1165         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1166                 return 0;
1167
1168         /* The backer wishes to know when pages are first written to? */
1169         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1170                 return 1;
1171
1172         /* The open routine did something to the protections already? */
1173         if (pgprot_val(vma->vm_page_prot) !=
1174             pgprot_val(vm_get_page_prot(vm_flags)))
1175                 return 0;
1176
1177         /* Specialty mapping? */
1178         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1179                 return 0;
1180
1181         /* Can the mapping track the dirty pages? */
1182         return vma->vm_file && vma->vm_file->f_mapping &&
1183                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1184 }
1185
1186 /*
1187  * We account for memory if it's a private writeable mapping,
1188  * not hugepages and VM_NORESERVE wasn't set.
1189  */
1190 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1191 {
1192         /*
1193          * hugetlb has its own accounting separate from the core VM
1194          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1195          */
1196         if (file && is_file_hugepages(file))
1197                 return 0;
1198
1199         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1200 }
1201
1202 unsigned long mmap_region(struct file *file, unsigned long addr,
1203                           unsigned long len, unsigned long flags,
1204                           vm_flags_t vm_flags, unsigned long pgoff)
1205 {
1206         struct mm_struct *mm = current->mm;
1207         struct vm_area_struct *vma, *prev;
1208         int error;
1209         struct rb_node **rb_link, *rb_parent;
1210         unsigned long charged = 0;
1211
1212         /* Clear old maps */
1213         error = -ENOMEM;
1214 munmap_back:
1215         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1216         if (vma && vma->vm_start < addr + len) {
1217                 if (do_munmap(mm, addr, len))
1218                         return -ENOMEM;
1219                 goto munmap_back;
1220         }
1221
1222         /* Check against address space limit. */
1223         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1224                 return -ENOMEM;
1225
1226         /*
1227          * Set 'VM_NORESERVE' if we should not account for the
1228          * memory use of this mapping.
1229          */
1230         if ((flags & MAP_NORESERVE)) {
1231                 /* We honor MAP_NORESERVE if allowed to overcommit */
1232                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1233                         vm_flags |= VM_NORESERVE;
1234
1235                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1236                 if (file && is_file_hugepages(file))
1237                         vm_flags |= VM_NORESERVE;
1238         }
1239
1240         /*
1241          * Private writable mapping: check memory availability
1242          */
1243         if (accountable_mapping(file, vm_flags)) {
1244                 charged = len >> PAGE_SHIFT;
1245                 if (security_vm_enough_memory(charged))
1246                         return -ENOMEM;
1247                 vm_flags |= VM_ACCOUNT;
1248         }
1249
1250         /*
1251          * Can we just expand an old mapping?
1252          */
1253         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1254         if (vma)
1255                 goto out;
1256
1257         /*
1258          * Determine the object being mapped and call the appropriate
1259          * specific mapper. the address has already been validated, but
1260          * not unmapped, but the maps are removed from the list.
1261          */
1262         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1263         if (!vma) {
1264                 error = -ENOMEM;
1265                 goto unacct_error;
1266         }
1267
1268         vma->vm_mm = mm;
1269         vma->vm_start = addr;
1270         vma->vm_end = addr + len;
1271         vma->vm_flags = vm_flags;
1272         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1273         vma->vm_pgoff = pgoff;
1274         INIT_LIST_HEAD(&vma->anon_vma_chain);
1275
1276         if (file) {
1277                 if (vm_flags & VM_DENYWRITE) {
1278                         error = deny_write_access(file);
1279                         if (error)
1280                                 goto free_vma;
1281                 }
1282                 if (vm_flags & VM_SHARED) {
1283                         error = mapping_map_writable(file->f_mapping);
1284                         if (error)
1285                                 goto allow_write_and_free_vma;
1286                 }
1287
1288                 /* ->mmap() can change vma->vm_file, but must guarantee that
1289                  * vma_link() below can deny write-access if VM_DENYWRITE is set
1290                  * and map writably if VM_SHARED is set. This usually means the
1291                  * new file must not have been exposed to user-space, yet.
1292                  */
1293                 vma->vm_file = get_file(file);
1294                 error = file->f_op->mmap(file, vma);
1295                 if (error)
1296                         goto unmap_and_free_vma;
1297                 if (vm_flags & VM_EXECUTABLE)
1298                         added_exe_file_vma(mm);
1299
1300                 /* Can addr have changed??
1301                  *
1302                  * Answer: Yes, several device drivers can do it in their
1303                  *         f_op->mmap method. -DaveM
1304                  */
1305                 addr = vma->vm_start;
1306                 pgoff = vma->vm_pgoff;
1307                 vm_flags = vma->vm_flags;
1308         } else if (vm_flags & VM_SHARED) {
1309                 error = shmem_zero_setup(vma);
1310                 if (error)
1311                         goto free_vma;
1312         }
1313
1314         if (vma_wants_writenotify(vma)) {
1315                 pgprot_t pprot = vma->vm_page_prot;
1316
1317                 /* Can vma->vm_page_prot have changed??
1318                  *
1319                  * Answer: Yes, drivers may have changed it in their
1320                  *         f_op->mmap method.
1321                  *
1322                  * Ensures that vmas marked as uncached stay that way.
1323                  */
1324                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1325                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1326                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1327         }
1328
1329         vma_link(mm, vma, prev, rb_link, rb_parent);
1330         /* Once vma denies write, undo our temporary denial count */
1331         if (file) {
1332                 if (vm_flags & VM_SHARED)
1333                         mapping_unmap_writable(file->f_mapping);
1334                 if (vm_flags & VM_DENYWRITE)
1335                         allow_write_access(file);
1336         }
1337         file = vma->vm_file;
1338 out:
1339         perf_event_mmap(vma);
1340
1341         mm->total_vm += len >> PAGE_SHIFT;
1342         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1343         if (vm_flags & VM_LOCKED) {
1344                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1345                         mm->locked_vm += (len >> PAGE_SHIFT);
1346         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1347                 make_pages_present(addr, addr + len);
1348         return addr;
1349
1350 unmap_and_free_vma:
1351         vma->vm_file = NULL;
1352         fput(file);
1353
1354         /* Undo any partial mapping done by a device driver. */
1355         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1356         charged = 0;
1357         if (vm_flags & VM_SHARED)
1358                 mapping_unmap_writable(file->f_mapping);
1359 allow_write_and_free_vma:
1360         if (vm_flags & VM_DENYWRITE)
1361                 allow_write_access(file);
1362 free_vma:
1363         kmem_cache_free(vm_area_cachep, vma);
1364 unacct_error:
1365         if (charged)
1366                 vm_unacct_memory(charged);
1367         return error;
1368 }
1369
1370 /* Get an address range which is currently unmapped.
1371  * For shmat() with addr=0.
1372  *
1373  * Ugly calling convention alert:
1374  * Return value with the low bits set means error value,
1375  * ie
1376  *      if (ret & ~PAGE_MASK)
1377  *              error = ret;
1378  *
1379  * This function "knows" that -ENOMEM has the bits set.
1380  */
1381 #ifndef HAVE_ARCH_UNMAPPED_AREA
1382 unsigned long
1383 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1384                 unsigned long len, unsigned long pgoff, unsigned long flags)
1385 {
1386         struct mm_struct *mm = current->mm;
1387         struct vm_area_struct *vma, *prev;
1388         unsigned long start_addr, vm_start, prev_end;
1389
1390         if (len > TASK_SIZE - mmap_min_addr)
1391                 return -ENOMEM;
1392
1393         if (flags & MAP_FIXED)
1394                 return addr;
1395
1396         if (addr) {
1397                 addr = PAGE_ALIGN(addr);
1398                 vma = find_vma_prev(mm, addr, &prev);
1399                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1400                     (!vma || addr + len <= vm_start_gap(vma)) &&
1401                     (!prev || addr >= vm_end_gap(prev)))
1402                         return addr;
1403         }
1404         if (len > mm->cached_hole_size) {
1405                 start_addr = addr = mm->free_area_cache;
1406         } else {
1407                 start_addr = addr = TASK_UNMAPPED_BASE;
1408                 mm->cached_hole_size = 0;
1409         }
1410
1411 full_search:
1412         for (vma = find_vma_prev(mm, addr, &prev); ; prev = vma,
1413                                                 vma = vma->vm_next) {
1414                 if (prev) {
1415                         prev_end = vm_end_gap(prev);
1416                         if (addr < prev_end) {
1417                                 addr = prev_end;
1418                                 /* If vma already violates gap, forget it */
1419                                 if (vma && addr > vma->vm_start)
1420                                         addr = vma->vm_start;
1421                         }
1422                 }
1423                 /* At this point:  (!vma || addr < vma->vm_end). */
1424                 if (TASK_SIZE - len < addr) {
1425                         /*
1426                          * Start a new search - just in case we missed
1427                          * some holes.
1428                          */
1429                         if (start_addr != TASK_UNMAPPED_BASE) {
1430                                 addr = TASK_UNMAPPED_BASE;
1431                                 start_addr = addr;
1432                                 mm->cached_hole_size = 0;
1433                                 goto full_search;
1434                         }
1435                         return -ENOMEM;
1436                 }
1437                 vm_start = vma ? vm_start_gap(vma) : TASK_SIZE;
1438                 if (addr + len <= vm_start) {
1439                         /*
1440                          * Remember the place where we stopped the search:
1441                          */
1442                         mm->free_area_cache = addr + len;
1443                         return addr;
1444                 }
1445                 if (addr + mm->cached_hole_size < vm_start)
1446                         mm->cached_hole_size = vm_start - addr;
1447         }
1448 }
1449 #endif  
1450
1451 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1452 {
1453         /*
1454          * Is this a new hole at the lowest possible address?
1455          */
1456         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1457                 mm->free_area_cache = addr;
1458                 mm->cached_hole_size = ~0UL;
1459         }
1460 }
1461
1462 /*
1463  * This mmap-allocator allocates new areas top-down from below the
1464  * stack's low limit (the base):
1465  */
1466 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1467 unsigned long
1468 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1469                           const unsigned long len, const unsigned long pgoff,
1470                           const unsigned long flags)
1471 {
1472         struct vm_area_struct *vma, *prev;
1473         struct mm_struct *mm = current->mm;
1474         unsigned long addr = addr0;
1475         unsigned long vm_start, prev_end;
1476         unsigned long low_limit = max(PAGE_SIZE, mmap_min_addr);
1477
1478         /* requested length too big for entire address space */
1479         if (len > TASK_SIZE - mmap_min_addr)
1480                 return -ENOMEM;
1481
1482         if (flags & MAP_FIXED)
1483                 return addr;
1484
1485         /* requesting a specific address */
1486         if (addr) {
1487                 addr = PAGE_ALIGN(addr);
1488                 vma = find_vma_prev(mm, addr, &prev);
1489                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1490                                 (!vma || addr + len <= vm_start_gap(vma)) &&
1491                                 (!prev || addr >= vm_end_gap(prev)))
1492                         return addr;
1493         }
1494
1495         /* check if free_area_cache is useful for us */
1496         if (len <= mm->cached_hole_size) {
1497                 mm->cached_hole_size = 0;
1498                 mm->free_area_cache = mm->mmap_base;
1499         }
1500
1501         /* either no address requested or can't fit in requested address hole */
1502         addr = mm->free_area_cache;
1503
1504         /* make sure it can fit in the remaining address space */
1505         if (addr >= low_limit + len) {
1506                 vma = find_vma_prev(mm, addr-len, &prev);
1507                 if ((!vma || addr <= vm_start_gap(vma)) &&
1508                     (!prev || addr-len >= vm_end_gap(prev)))
1509                         /* remember the address as a hint for next time */
1510                         return (mm->free_area_cache = addr-len);
1511         }
1512
1513         if (mm->mmap_base < low_limit + len)
1514                 goto bottomup;
1515
1516         addr = mm->mmap_base-len;
1517
1518         do {
1519                 /*
1520                  * Lookup failure means no vma is above this address,
1521                  * else if new region fits below vma->vm_start,
1522                  * return with success:
1523                  */
1524                 vma = find_vma_prev(mm, addr, &prev);
1525                 vm_start = vma ? vm_start_gap(vma) : mm->mmap_base;
1526                 prev_end = prev ? vm_end_gap(prev) : low_limit;
1527
1528                 if (addr + len <= vm_start && addr >= prev_end)
1529                         /* remember the address as a hint for next time */
1530                         return (mm->free_area_cache = addr);
1531
1532                 /* remember the largest hole we saw so far */
1533                 if (addr + mm->cached_hole_size < vm_start)
1534                         mm->cached_hole_size = vm_start - addr;
1535
1536                 /* try just below the current vma->vm_start */
1537                 addr = vm_start - len;
1538         } while (vm_start >= low_limit + len);
1539
1540 bottomup:
1541         /*
1542          * A failed mmap() very likely causes application failure,
1543          * so fall back to the bottom-up function here. This scenario
1544          * can happen with large stack limits and large mmap()
1545          * allocations.
1546          */
1547         mm->cached_hole_size = ~0UL;
1548         mm->free_area_cache = TASK_UNMAPPED_BASE;
1549         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1550         /*
1551          * Restore the topdown base:
1552          */
1553         mm->free_area_cache = mm->mmap_base;
1554         mm->cached_hole_size = ~0UL;
1555
1556         return addr;
1557 }
1558 #endif
1559
1560 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1561 {
1562         /*
1563          * Is this a new hole at the highest possible address?
1564          */
1565         if (addr > mm->free_area_cache)
1566                 mm->free_area_cache = addr;
1567
1568         /* dont allow allocations above current base */
1569         if (mm->free_area_cache > mm->mmap_base)
1570                 mm->free_area_cache = mm->mmap_base;
1571 }
1572
1573 unsigned long
1574 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1575                 unsigned long pgoff, unsigned long flags)
1576 {
1577         unsigned long (*get_area)(struct file *, unsigned long,
1578                                   unsigned long, unsigned long, unsigned long);
1579
1580         unsigned long error = arch_mmap_check(addr, len, flags);
1581         if (error)
1582                 return error;
1583
1584         /* Careful about overflows.. */
1585         if (len > TASK_SIZE)
1586                 return -ENOMEM;
1587
1588         get_area = current->mm->get_unmapped_area;
1589         if (file && file->f_op && file->f_op->get_unmapped_area)
1590                 get_area = file->f_op->get_unmapped_area;
1591         addr = get_area(file, addr, len, pgoff, flags);
1592         if (IS_ERR_VALUE(addr))
1593                 return addr;
1594
1595         if (addr > TASK_SIZE - len)
1596                 return -ENOMEM;
1597         if (addr & ~PAGE_MASK)
1598                 return -EINVAL;
1599
1600         return arch_rebalance_pgtables(addr, len);
1601 }
1602
1603 EXPORT_SYMBOL(get_unmapped_area);
1604
1605 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1606 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1607 {
1608         struct vm_area_struct *vma = NULL;
1609
1610         if (mm) {
1611                 /* Check the cache first. */
1612                 /* (Cache hit rate is typically around 35%.) */
1613                 vma = ACCESS_ONCE(mm->mmap_cache);
1614                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1615                         struct rb_node * rb_node;
1616
1617                         rb_node = mm->mm_rb.rb_node;
1618                         vma = NULL;
1619
1620                         while (rb_node) {
1621                                 struct vm_area_struct * vma_tmp;
1622
1623                                 vma_tmp = rb_entry(rb_node,
1624                                                 struct vm_area_struct, vm_rb);
1625
1626                                 if (vma_tmp->vm_end > addr) {
1627                                         vma = vma_tmp;
1628                                         if (vma_tmp->vm_start <= addr)
1629                                                 break;
1630                                         rb_node = rb_node->rb_left;
1631                                 } else
1632                                         rb_node = rb_node->rb_right;
1633                         }
1634                         if (vma)
1635                                 mm->mmap_cache = vma;
1636                 }
1637         }
1638         return vma;
1639 }
1640
1641 EXPORT_SYMBOL(find_vma);
1642
1643 /*
1644  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1645  */
1646 struct vm_area_struct *
1647 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1648                         struct vm_area_struct **pprev)
1649 {
1650         struct vm_area_struct *vma;
1651
1652         vma = find_vma(mm, addr);
1653         if (vma) {
1654                 *pprev = vma->vm_prev;
1655         } else {
1656                 struct rb_node *rb_node = mm->mm_rb.rb_node;
1657                 *pprev = NULL;
1658                 while (rb_node) {
1659                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1660                         rb_node = rb_node->rb_right;
1661                 }
1662         }
1663         return vma;
1664 }
1665
1666 /*
1667  * Verify that the stack growth is acceptable and
1668  * update accounting. This is shared with both the
1669  * grow-up and grow-down cases.
1670  */
1671 static int acct_stack_growth(struct vm_area_struct *vma,
1672                              unsigned long size, unsigned long grow)
1673 {
1674         struct mm_struct *mm = vma->vm_mm;
1675         struct rlimit *rlim = current->signal->rlim;
1676         unsigned long new_start;
1677
1678         /* address space limit tests */
1679         if (!may_expand_vm(mm, grow))
1680                 return -ENOMEM;
1681
1682         /* Stack limit test */
1683         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1684                 return -ENOMEM;
1685
1686         /* mlock limit tests */
1687         if (vma->vm_flags & VM_LOCKED) {
1688                 unsigned long locked;
1689                 unsigned long limit;
1690                 locked = mm->locked_vm + grow;
1691                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1692                 limit >>= PAGE_SHIFT;
1693                 if (locked > limit && !capable(CAP_IPC_LOCK))
1694                         return -ENOMEM;
1695         }
1696
1697         /* Check to ensure the stack will not grow into a hugetlb-only region */
1698         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1699                         vma->vm_end - size;
1700         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1701                 return -EFAULT;
1702
1703         /*
1704          * Overcommit..  This must be the final test, as it will
1705          * update security statistics.
1706          */
1707         if (security_vm_enough_memory_mm(mm, grow))
1708                 return -ENOMEM;
1709
1710         /* Ok, everything looks good - let it rip */
1711         mm->total_vm += grow;
1712         if (vma->vm_flags & VM_LOCKED)
1713                 mm->locked_vm += grow;
1714         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1715         return 0;
1716 }
1717
1718 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1719 /*
1720  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1721  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1722  */
1723 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1724 {
1725         struct vm_area_struct *next;
1726         unsigned long gap_addr;
1727         int error = 0;
1728
1729         if (!(vma->vm_flags & VM_GROWSUP))
1730                 return -EFAULT;
1731
1732         /* Guard against exceeding limits of the address space. */
1733         address &= PAGE_MASK;
1734         if (address >= (TASK_SIZE & PAGE_MASK))
1735                 return -ENOMEM;
1736         address += PAGE_SIZE;
1737
1738         /* Enforce stack_guard_gap */
1739         gap_addr = address + stack_guard_gap;
1740
1741         /* Guard against overflow */
1742         if (gap_addr < address || gap_addr > TASK_SIZE)
1743                 gap_addr = TASK_SIZE;
1744
1745         next = vma->vm_next;
1746         if (next && next->vm_start < gap_addr &&
1747                         (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
1748                 if (!(next->vm_flags & VM_GROWSUP))
1749                         return -ENOMEM;
1750                 /* Check that both stack segments have the same anon_vma? */
1751         }
1752
1753         /* We must make sure the anon_vma is allocated. */
1754         if (unlikely(anon_vma_prepare(vma)))
1755                 return -ENOMEM;
1756
1757         /*
1758          * vma->vm_start/vm_end cannot change under us because the caller
1759          * is required to hold the mmap_sem in read mode.  We need the
1760          * anon_vma lock to serialize against concurrent expand_stacks.
1761          */
1762         vma_lock_anon_vma(vma);
1763
1764         /* Somebody else might have raced and expanded it already */
1765         if (address > vma->vm_end) {
1766                 unsigned long size, grow;
1767
1768                 size = address - vma->vm_start;
1769                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1770
1771                 error = -ENOMEM;
1772                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1773                         error = acct_stack_growth(vma, size, grow);
1774                         if (!error) {
1775                                 vma->vm_end = address;
1776                                 perf_event_mmap(vma);
1777                         }
1778                 }
1779         }
1780         vma_unlock_anon_vma(vma);
1781         khugepaged_enter_vma_merge(vma, vma->vm_flags);
1782         return error;
1783 }
1784 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1785
1786 /*
1787  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1788  */
1789 int expand_downwards(struct vm_area_struct *vma,
1790                                    unsigned long address)
1791 {
1792         struct vm_area_struct *prev;
1793         unsigned long gap_addr;
1794         int error;
1795
1796         address &= PAGE_MASK;
1797         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1798         if (error)
1799                 return error;
1800
1801         /* Enforce stack_guard_gap */
1802         gap_addr = address - stack_guard_gap;
1803         if (gap_addr > address)
1804                 return -ENOMEM;
1805         prev = vma->vm_prev;
1806         if (prev && prev->vm_end > gap_addr &&
1807                         (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
1808                 if (!(prev->vm_flags & VM_GROWSDOWN))
1809                         return -ENOMEM;
1810                 /* Check that both stack segments have the same anon_vma? */
1811         }
1812
1813         /* We must make sure the anon_vma is allocated. */
1814         if (unlikely(anon_vma_prepare(vma)))
1815                 return -ENOMEM;
1816
1817         /*
1818          * vma->vm_start/vm_end cannot change under us because the caller
1819          * is required to hold the mmap_sem in read mode.  We need the
1820          * anon_vma lock to serialize against concurrent expand_stacks.
1821          */
1822         vma_lock_anon_vma(vma);
1823
1824         /* Somebody else might have raced and expanded it already */
1825         if (address < vma->vm_start) {
1826                 unsigned long size, grow;
1827
1828                 size = vma->vm_end - address;
1829                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1830
1831                 error = -ENOMEM;
1832                 if (grow <= vma->vm_pgoff) {
1833                         error = acct_stack_growth(vma, size, grow);
1834                         if (!error) {
1835                                 vma->vm_start = address;
1836                                 vma->vm_pgoff -= grow;
1837                                 perf_event_mmap(vma);
1838                         }
1839                 }
1840         }
1841         vma_unlock_anon_vma(vma);
1842         khugepaged_enter_vma_merge(vma, vma->vm_flags);
1843         return error;
1844 }
1845
1846 /* enforced gap between the expanding stack and other mappings. */
1847 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
1848
1849 static int __init cmdline_parse_stack_guard_gap(char *p)
1850 {
1851         unsigned long val;
1852         char *endptr;
1853
1854         val = simple_strtoul(p, &endptr, 10);
1855         if (!*endptr)
1856                 stack_guard_gap = val << PAGE_SHIFT;
1857
1858         return 0;
1859 }
1860 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
1861
1862 #ifdef CONFIG_STACK_GROWSUP
1863 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1864 {
1865         return expand_upwards(vma, address);
1866 }
1867
1868 struct vm_area_struct *
1869 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1870 {
1871         struct vm_area_struct *vma, *prev;
1872
1873         addr &= PAGE_MASK;
1874         vma = find_vma_prev(mm, addr, &prev);
1875         if (vma && (vma->vm_start <= addr))
1876                 return vma;
1877         if (!prev || expand_stack(prev, addr))
1878                 return NULL;
1879         if (prev->vm_flags & VM_LOCKED) {
1880                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1881         }
1882         return prev;
1883 }
1884 #else
1885 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1886 {
1887         return expand_downwards(vma, address);
1888 }
1889
1890 struct vm_area_struct *
1891 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1892 {
1893         struct vm_area_struct * vma;
1894         unsigned long start;
1895
1896         addr &= PAGE_MASK;
1897         vma = find_vma(mm,addr);
1898         if (!vma)
1899                 return NULL;
1900         if (vma->vm_start <= addr)
1901                 return vma;
1902         if (!(vma->vm_flags & VM_GROWSDOWN))
1903                 return NULL;
1904         start = vma->vm_start;
1905         if (expand_stack(vma, addr))
1906                 return NULL;
1907         if (vma->vm_flags & VM_LOCKED) {
1908                 mlock_vma_pages_range(vma, addr, start);
1909         }
1910         return vma;
1911 }
1912 #endif
1913
1914 /*
1915  * Ok - we have the memory areas we should free on the vma list,
1916  * so release them, and do the vma updates.
1917  *
1918  * Called with the mm semaphore held.
1919  */
1920 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1921 {
1922         /* Update high watermark before we lower total_vm */
1923         update_hiwater_vm(mm);
1924         do {
1925                 long nrpages = vma_pages(vma);
1926
1927                 mm->total_vm -= nrpages;
1928                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1929                 vma = remove_vma(vma);
1930         } while (vma);
1931         validate_mm(mm);
1932 }
1933
1934 /*
1935  * Get rid of page table information in the indicated region.
1936  *
1937  * Called with the mm semaphore held.
1938  */
1939 static void unmap_region(struct mm_struct *mm,
1940                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1941                 unsigned long start, unsigned long end)
1942 {
1943         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1944         struct mmu_gather tlb;
1945         unsigned long nr_accounted = 0;
1946
1947         lru_add_drain();
1948         tlb_gather_mmu(&tlb, mm, 0);
1949         update_hiwater_rss(mm);
1950         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1951         vm_unacct_memory(nr_accounted);
1952         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1953                                  next ? next->vm_start : 0);
1954         tlb_finish_mmu(&tlb, start, end);
1955 }
1956
1957 /*
1958  * Create a list of vma's touched by the unmap, removing them from the mm's
1959  * vma list as we go..
1960  */
1961 static void
1962 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1963         struct vm_area_struct *prev, unsigned long end)
1964 {
1965         struct vm_area_struct **insertion_point;
1966         struct vm_area_struct *tail_vma = NULL;
1967         unsigned long addr;
1968
1969         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1970         vma->vm_prev = NULL;
1971         do {
1972                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1973                 mm->map_count--;
1974                 tail_vma = vma;
1975                 vma = vma->vm_next;
1976         } while (vma && vma->vm_start < end);
1977         *insertion_point = vma;
1978         if (vma)
1979                 vma->vm_prev = prev;
1980         tail_vma->vm_next = NULL;
1981         if (mm->unmap_area == arch_unmap_area)
1982                 addr = prev ? prev->vm_end : mm->mmap_base;
1983         else
1984                 addr = vma ?  vma->vm_start : mm->mmap_base;
1985         mm->unmap_area(mm, addr);
1986         mm->mmap_cache = NULL;          /* Kill the cache. */
1987 }
1988
1989 /*
1990  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1991  * munmap path where it doesn't make sense to fail.
1992  */
1993 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1994               unsigned long addr, int new_below)
1995 {
1996         struct mempolicy *pol;
1997         struct vm_area_struct *new;
1998         int err = -ENOMEM;
1999
2000         if (is_vm_hugetlb_page(vma) && (addr &
2001                                         ~(huge_page_mask(hstate_vma(vma)))))
2002                 return -EINVAL;
2003
2004         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2005         if (!new)
2006                 goto out_err;
2007
2008         /* most fields are the same, copy all, and then fixup */
2009         *new = *vma;
2010
2011         INIT_LIST_HEAD(&new->anon_vma_chain);
2012
2013         if (new_below)
2014                 new->vm_end = addr;
2015         else {
2016                 new->vm_start = addr;
2017                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2018         }
2019
2020         pol = mpol_dup(vma_policy(vma));
2021         if (IS_ERR(pol)) {
2022                 err = PTR_ERR(pol);
2023                 goto out_free_vma;
2024         }
2025         vma_set_policy(new, pol);
2026
2027         if (anon_vma_clone(new, vma))
2028                 goto out_free_mpol;
2029
2030         if (new->vm_file) {
2031                 get_file(new->vm_file);
2032                 if (vma->vm_flags & VM_EXECUTABLE)
2033                         added_exe_file_vma(mm);
2034         }
2035
2036         if (new->vm_ops && new->vm_ops->open)
2037                 new->vm_ops->open(new);
2038
2039         if (new_below)
2040                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2041                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2042         else
2043                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2044
2045         /* Success. */
2046         if (!err)
2047                 return 0;
2048
2049         /* Clean everything up if vma_adjust failed. */
2050         if (new->vm_ops && new->vm_ops->close)
2051                 new->vm_ops->close(new);
2052         if (new->vm_file) {
2053                 if (vma->vm_flags & VM_EXECUTABLE)
2054                         removed_exe_file_vma(mm);
2055                 fput(new->vm_file);
2056         }
2057         unlink_anon_vmas(new);
2058  out_free_mpol:
2059         mpol_put(pol);
2060  out_free_vma:
2061         kmem_cache_free(vm_area_cachep, new);
2062  out_err:
2063         return err;
2064 }
2065
2066 /*
2067  * Split a vma into two pieces at address 'addr', a new vma is allocated
2068  * either for the first part or the tail.
2069  */
2070 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2071               unsigned long addr, int new_below)
2072 {
2073         if (mm->map_count >= sysctl_max_map_count)
2074                 return -ENOMEM;
2075
2076         return __split_vma(mm, vma, addr, new_below);
2077 }
2078
2079 /* Munmap is split into 2 main parts -- this part which finds
2080  * what needs doing, and the areas themselves, which do the
2081  * work.  This now handles partial unmappings.
2082  * Jeremy Fitzhardinge <jeremy@goop.org>
2083  */
2084 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2085 {
2086         unsigned long end;
2087         struct vm_area_struct *vma, *prev, *last;
2088
2089         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2090                 return -EINVAL;
2091
2092         if ((len = PAGE_ALIGN(len)) == 0)
2093                 return -EINVAL;
2094
2095         /* Find the first overlapping VMA */
2096         vma = find_vma(mm, start);
2097         if (!vma)
2098                 return 0;
2099         prev = vma->vm_prev;
2100         /* we have  start < vma->vm_end  */
2101
2102         /* if it doesn't overlap, we have nothing.. */
2103         end = start + len;
2104         if (vma->vm_start >= end)
2105                 return 0;
2106
2107         /*
2108          * If we need to split any vma, do it now to save pain later.
2109          *
2110          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2111          * unmapped vm_area_struct will remain in use: so lower split_vma
2112          * places tmp vma above, and higher split_vma places tmp vma below.
2113          */
2114         if (start > vma->vm_start) {
2115                 int error;
2116
2117                 /*
2118                  * Make sure that map_count on return from munmap() will
2119                  * not exceed its limit; but let map_count go just above
2120                  * its limit temporarily, to help free resources as expected.
2121                  */
2122                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2123                         return -ENOMEM;
2124
2125                 error = __split_vma(mm, vma, start, 0);
2126                 if (error)
2127                         return error;
2128                 prev = vma;
2129         }
2130
2131         /* Does it split the last one? */
2132         last = find_vma(mm, end);
2133         if (last && end > last->vm_start) {
2134                 int error = __split_vma(mm, last, end, 1);
2135                 if (error)
2136                         return error;
2137         }
2138         vma = prev? prev->vm_next: mm->mmap;
2139
2140         /*
2141          * unlock any mlock()ed ranges before detaching vmas
2142          */
2143         if (mm->locked_vm) {
2144                 struct vm_area_struct *tmp = vma;
2145                 while (tmp && tmp->vm_start < end) {
2146                         if (tmp->vm_flags & VM_LOCKED) {
2147                                 mm->locked_vm -= vma_pages(tmp);
2148                                 munlock_vma_pages_all(tmp);
2149                         }
2150                         tmp = tmp->vm_next;
2151                 }
2152         }
2153
2154         /*
2155          * Remove the vma's, and unmap the actual pages
2156          */
2157         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2158         unmap_region(mm, vma, prev, start, end);
2159
2160         /* Fix up all other VM information */
2161         remove_vma_list(mm, vma);
2162
2163         return 0;
2164 }
2165
2166 EXPORT_SYMBOL(do_munmap);
2167
2168 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2169 {
2170         int ret;
2171         struct mm_struct *mm = current->mm;
2172
2173         profile_munmap(addr);
2174
2175         down_write(&mm->mmap_sem);
2176         ret = do_munmap(mm, addr, len);
2177         up_write(&mm->mmap_sem);
2178         return ret;
2179 }
2180
2181 static inline void verify_mm_writelocked(struct mm_struct *mm)
2182 {
2183 #ifdef CONFIG_DEBUG_VM
2184         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2185                 WARN_ON(1);
2186                 up_read(&mm->mmap_sem);
2187         }
2188 #endif
2189 }
2190
2191 /*
2192  *  this is really a simplified "do_mmap".  it only handles
2193  *  anonymous maps.  eventually we may be able to do some
2194  *  brk-specific accounting here.
2195  */
2196 unsigned long do_brk(unsigned long addr, unsigned long len)
2197 {
2198         struct mm_struct * mm = current->mm;
2199         struct vm_area_struct * vma, * prev;
2200         unsigned long flags;
2201         struct rb_node ** rb_link, * rb_parent;
2202         pgoff_t pgoff = addr >> PAGE_SHIFT;
2203         int error;
2204
2205         len = PAGE_ALIGN(len);
2206         if (!len)
2207                 return addr;
2208
2209         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2210         if (error)
2211                 return error;
2212
2213         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2214
2215         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2216         if (error & ~PAGE_MASK)
2217                 return error;
2218
2219         /*
2220          * mlock MCL_FUTURE?
2221          */
2222         if (mm->def_flags & VM_LOCKED) {
2223                 unsigned long locked, lock_limit;
2224                 locked = len >> PAGE_SHIFT;
2225                 locked += mm->locked_vm;
2226                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2227                 lock_limit >>= PAGE_SHIFT;
2228                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2229                         return -EAGAIN;
2230         }
2231
2232         /*
2233          * mm->mmap_sem is required to protect against another thread
2234          * changing the mappings in case we sleep.
2235          */
2236         verify_mm_writelocked(mm);
2237
2238         /*
2239          * Clear old maps.  this also does some error checking for us
2240          */
2241  munmap_back:
2242         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2243         if (vma && vma->vm_start < addr + len) {
2244                 if (do_munmap(mm, addr, len))
2245                         return -ENOMEM;
2246                 goto munmap_back;
2247         }
2248
2249         /* Check against address space limits *after* clearing old maps... */
2250         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2251                 return -ENOMEM;
2252
2253         if (mm->map_count > sysctl_max_map_count)
2254                 return -ENOMEM;
2255
2256         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2257                 return -ENOMEM;
2258
2259         /* Can we just expand an old private anonymous mapping? */
2260         vma = vma_merge(mm, prev, addr, addr + len, flags,
2261                                         NULL, NULL, pgoff, NULL);
2262         if (vma)
2263                 goto out;
2264
2265         /*
2266          * create a vma struct for an anonymous mapping
2267          */
2268         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2269         if (!vma) {
2270                 vm_unacct_memory(len >> PAGE_SHIFT);
2271                 return -ENOMEM;
2272         }
2273
2274         INIT_LIST_HEAD(&vma->anon_vma_chain);
2275         vma->vm_mm = mm;
2276         vma->vm_start = addr;
2277         vma->vm_end = addr + len;
2278         vma->vm_pgoff = pgoff;
2279         vma->vm_flags = flags;
2280         vma->vm_page_prot = vm_get_page_prot(flags);
2281         vma_link(mm, vma, prev, rb_link, rb_parent);
2282 out:
2283         perf_event_mmap(vma);
2284         mm->total_vm += len >> PAGE_SHIFT;
2285         if (flags & VM_LOCKED) {
2286                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2287                         mm->locked_vm += (len >> PAGE_SHIFT);
2288         }
2289         return addr;
2290 }
2291
2292 EXPORT_SYMBOL(do_brk);
2293
2294 /* Release all mmaps. */
2295 void exit_mmap(struct mm_struct *mm)
2296 {
2297         struct mmu_gather tlb;
2298         struct vm_area_struct *vma;
2299         unsigned long nr_accounted = 0;
2300         unsigned long end;
2301
2302         /* mm's last user has gone, and its about to be pulled down */
2303         mmu_notifier_release(mm);
2304
2305         if (mm->locked_vm) {
2306                 vma = mm->mmap;
2307                 while (vma) {
2308                         if (vma->vm_flags & VM_LOCKED)
2309                                 munlock_vma_pages_all(vma);
2310                         vma = vma->vm_next;
2311                 }
2312         }
2313
2314         arch_exit_mmap(mm);
2315
2316         vma = mm->mmap;
2317         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2318                 return;
2319
2320         lru_add_drain();
2321         flush_cache_mm(mm);
2322         tlb_gather_mmu(&tlb, mm, 1);
2323         /* update_hiwater_rss(mm) here? but nobody should be looking */
2324         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2325         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2326         vm_unacct_memory(nr_accounted);
2327
2328         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2329         tlb_finish_mmu(&tlb, 0, end);
2330
2331         /*
2332          * Walk the list again, actually closing and freeing it,
2333          * with preemption enabled, without holding any MM locks.
2334          */
2335         while (vma)
2336                 vma = remove_vma(vma);
2337
2338         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2339 }
2340
2341 /* Insert vm structure into process list sorted by address
2342  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2343  * then i_mmap_mutex is taken here.
2344  */
2345 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2346 {
2347         struct vm_area_struct * __vma, * prev;
2348         struct rb_node ** rb_link, * rb_parent;
2349
2350         /*
2351          * The vm_pgoff of a purely anonymous vma should be irrelevant
2352          * until its first write fault, when page's anon_vma and index
2353          * are set.  But now set the vm_pgoff it will almost certainly
2354          * end up with (unless mremap moves it elsewhere before that
2355          * first wfault), so /proc/pid/maps tells a consistent story.
2356          *
2357          * By setting it to reflect the virtual start address of the
2358          * vma, merges and splits can happen in a seamless way, just
2359          * using the existing file pgoff checks and manipulations.
2360          * Similarly in do_mmap_pgoff and in do_brk.
2361          */
2362         if (!vma->vm_file) {
2363                 BUG_ON(vma->anon_vma);
2364                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2365         }
2366         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2367         if (__vma && __vma->vm_start < vma->vm_end)
2368                 return -ENOMEM;
2369         if ((vma->vm_flags & VM_ACCOUNT) &&
2370              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2371                 return -ENOMEM;
2372         vma_link(mm, vma, prev, rb_link, rb_parent);
2373         return 0;
2374 }
2375
2376 /*
2377  * Copy the vma structure to a new location in the same mm,
2378  * prior to moving page table entries, to effect an mremap move.
2379  */
2380 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2381         unsigned long addr, unsigned long len, pgoff_t pgoff)
2382 {
2383         struct vm_area_struct *vma = *vmap;
2384         unsigned long vma_start = vma->vm_start;
2385         struct mm_struct *mm = vma->vm_mm;
2386         struct vm_area_struct *new_vma, *prev;
2387         struct rb_node **rb_link, *rb_parent;
2388         struct mempolicy *pol;
2389
2390         /*
2391          * If anonymous vma has not yet been faulted, update new pgoff
2392          * to match new location, to increase its chance of merging.
2393          */
2394         if (!vma->vm_file && !vma->anon_vma)
2395                 pgoff = addr >> PAGE_SHIFT;
2396
2397         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2398         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2399                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2400         if (new_vma) {
2401                 /*
2402                  * Source vma may have been merged into new_vma
2403                  */
2404                 if (vma_start >= new_vma->vm_start &&
2405                     vma_start < new_vma->vm_end)
2406                         *vmap = new_vma;
2407         } else {
2408                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2409                 if (new_vma) {
2410                         *new_vma = *vma;
2411                         pol = mpol_dup(vma_policy(vma));
2412                         if (IS_ERR(pol))
2413                                 goto out_free_vma;
2414                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2415                         if (anon_vma_clone(new_vma, vma))
2416                                 goto out_free_mempol;
2417                         vma_set_policy(new_vma, pol);
2418                         new_vma->vm_start = addr;
2419                         new_vma->vm_end = addr + len;
2420                         new_vma->vm_pgoff = pgoff;
2421                         if (new_vma->vm_file) {
2422                                 get_file(new_vma->vm_file);
2423                                 if (vma->vm_flags & VM_EXECUTABLE)
2424                                         added_exe_file_vma(mm);
2425                         }
2426                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2427                                 new_vma->vm_ops->open(new_vma);
2428                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2429                 }
2430         }
2431         return new_vma;
2432
2433  out_free_mempol:
2434         mpol_put(pol);
2435  out_free_vma:
2436         kmem_cache_free(vm_area_cachep, new_vma);
2437         return NULL;
2438 }
2439
2440 /*
2441  * Return true if the calling process may expand its vm space by the passed
2442  * number of pages
2443  */
2444 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2445 {
2446         unsigned long cur = mm->total_vm;       /* pages */
2447         unsigned long lim;
2448
2449         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2450
2451         if (cur + npages > lim)
2452                 return 0;
2453         return 1;
2454 }
2455
2456
2457 static int special_mapping_fault(struct vm_area_struct *vma,
2458                                 struct vm_fault *vmf)
2459 {
2460         pgoff_t pgoff;
2461         struct page **pages;
2462
2463         /*
2464          * special mappings have no vm_file, and in that case, the mm
2465          * uses vm_pgoff internally. So we have to subtract it from here.
2466          * We are allowed to do this because we are the mm; do not copy
2467          * this code into drivers!
2468          */
2469         pgoff = vmf->pgoff - vma->vm_pgoff;
2470
2471         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2472                 pgoff--;
2473
2474         if (*pages) {
2475                 struct page *page = *pages;
2476                 get_page(page);
2477                 vmf->page = page;
2478                 return 0;
2479         }
2480
2481         return VM_FAULT_SIGBUS;
2482 }
2483
2484 /*
2485  * Having a close hook prevents vma merging regardless of flags.
2486  */
2487 static void special_mapping_close(struct vm_area_struct *vma)
2488 {
2489 }
2490
2491 static const struct vm_operations_struct special_mapping_vmops = {
2492         .close = special_mapping_close,
2493         .fault = special_mapping_fault,
2494 };
2495
2496 /*
2497  * Called with mm->mmap_sem held for writing.
2498  * Insert a new vma covering the given region, with the given flags.
2499  * Its pages are supplied by the given array of struct page *.
2500  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2501  * The region past the last page supplied will always produce SIGBUS.
2502  * The array pointer and the pages it points to are assumed to stay alive
2503  * for as long as this mapping might exist.
2504  */
2505 int install_special_mapping(struct mm_struct *mm,
2506                             unsigned long addr, unsigned long len,
2507                             unsigned long vm_flags, struct page **pages)
2508 {
2509         int ret;
2510         struct vm_area_struct *vma;
2511
2512         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2513         if (unlikely(vma == NULL))
2514                 return -ENOMEM;
2515
2516         INIT_LIST_HEAD(&vma->anon_vma_chain);
2517         vma->vm_mm = mm;
2518         vma->vm_start = addr;
2519         vma->vm_end = addr + len;
2520
2521         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2522         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2523
2524         vma->vm_ops = &special_mapping_vmops;
2525         vma->vm_private_data = pages;
2526
2527         ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2528         if (ret)
2529                 goto out;
2530
2531         ret = insert_vm_struct(mm, vma);
2532         if (ret)
2533                 goto out;
2534
2535         mm->total_vm += len >> PAGE_SHIFT;
2536
2537         perf_event_mmap(vma);
2538
2539         return 0;
2540
2541 out:
2542         kmem_cache_free(vm_area_cachep, vma);
2543         return ret;
2544 }
2545
2546 static DEFINE_MUTEX(mm_all_locks_mutex);
2547
2548 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2549 {
2550         if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2551                 /*
2552                  * The LSB of head.next can't change from under us
2553                  * because we hold the mm_all_locks_mutex.
2554                  */
2555                 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2556                 /*
2557                  * We can safely modify head.next after taking the
2558                  * anon_vma->root->mutex. If some other vma in this mm shares
2559                  * the same anon_vma we won't take it again.
2560                  *
2561                  * No need of atomic instructions here, head.next
2562                  * can't change from under us thanks to the
2563                  * anon_vma->root->mutex.
2564                  */
2565                 if (__test_and_set_bit(0, (unsigned long *)
2566                                        &anon_vma->root->head.next))
2567                         BUG();
2568         }
2569 }
2570
2571 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2572 {
2573         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2574                 /*
2575                  * AS_MM_ALL_LOCKS can't change from under us because
2576                  * we hold the mm_all_locks_mutex.
2577                  *
2578                  * Operations on ->flags have to be atomic because
2579                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2580                  * mm_all_locks_mutex, there may be other cpus
2581                  * changing other bitflags in parallel to us.
2582                  */
2583                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2584                         BUG();
2585                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2586         }
2587 }
2588
2589 /*
2590  * This operation locks against the VM for all pte/vma/mm related
2591  * operations that could ever happen on a certain mm. This includes
2592  * vmtruncate, try_to_unmap, and all page faults.
2593  *
2594  * The caller must take the mmap_sem in write mode before calling
2595  * mm_take_all_locks(). The caller isn't allowed to release the
2596  * mmap_sem until mm_drop_all_locks() returns.
2597  *
2598  * mmap_sem in write mode is required in order to block all operations
2599  * that could modify pagetables and free pages without need of
2600  * altering the vma layout (for example populate_range() with
2601  * nonlinear vmas). It's also needed in write mode to avoid new
2602  * anon_vmas to be associated with existing vmas.
2603  *
2604  * A single task can't take more than one mm_take_all_locks() in a row
2605  * or it would deadlock.
2606  *
2607  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2608  * mapping->flags avoid to take the same lock twice, if more than one
2609  * vma in this mm is backed by the same anon_vma or address_space.
2610  *
2611  * We can take all the locks in random order because the VM code
2612  * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2613  * takes more than one of them in a row. Secondly we're protected
2614  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2615  *
2616  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2617  * that may have to take thousand of locks.
2618  *
2619  * mm_take_all_locks() can fail if it's interrupted by signals.
2620  */
2621 int mm_take_all_locks(struct mm_struct *mm)
2622 {
2623         struct vm_area_struct *vma;
2624         struct anon_vma_chain *avc;
2625
2626         BUG_ON(down_read_trylock(&mm->mmap_sem));
2627
2628         mutex_lock(&mm_all_locks_mutex);
2629
2630         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2631                 if (signal_pending(current))
2632                         goto out_unlock;
2633                 if (vma->vm_file && vma->vm_file->f_mapping)
2634                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2635         }
2636
2637         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2638                 if (signal_pending(current))
2639                         goto out_unlock;
2640                 if (vma->anon_vma)
2641                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2642                                 vm_lock_anon_vma(mm, avc->anon_vma);
2643         }
2644
2645         return 0;
2646
2647 out_unlock:
2648         mm_drop_all_locks(mm);
2649         return -EINTR;
2650 }
2651
2652 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2653 {
2654         if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2655                 /*
2656                  * The LSB of head.next can't change to 0 from under
2657                  * us because we hold the mm_all_locks_mutex.
2658                  *
2659                  * We must however clear the bitflag before unlocking
2660                  * the vma so the users using the anon_vma->head will
2661                  * never see our bitflag.
2662                  *
2663                  * No need of atomic instructions here, head.next
2664                  * can't change from under us until we release the
2665                  * anon_vma->root->mutex.
2666                  */
2667                 if (!__test_and_clear_bit(0, (unsigned long *)
2668                                           &anon_vma->root->head.next))
2669                         BUG();
2670                 anon_vma_unlock(anon_vma);
2671         }
2672 }
2673
2674 static void vm_unlock_mapping(struct address_space *mapping)
2675 {
2676         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2677                 /*
2678                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2679                  * because we hold the mm_all_locks_mutex.
2680                  */
2681                 mutex_unlock(&mapping->i_mmap_mutex);
2682                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2683                                         &mapping->flags))
2684                         BUG();
2685         }
2686 }
2687
2688 /*
2689  * The mmap_sem cannot be released by the caller until
2690  * mm_drop_all_locks() returns.
2691  */
2692 void mm_drop_all_locks(struct mm_struct *mm)
2693 {
2694         struct vm_area_struct *vma;
2695         struct anon_vma_chain *avc;
2696
2697         BUG_ON(down_read_trylock(&mm->mmap_sem));
2698         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2699
2700         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2701                 if (vma->anon_vma)
2702                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2703                                 vm_unlock_anon_vma(avc->anon_vma);
2704                 if (vma->vm_file && vma->vm_file->f_mapping)
2705                         vm_unlock_mapping(vma->vm_file->f_mapping);
2706         }
2707
2708         mutex_unlock(&mm_all_locks_mutex);
2709 }
2710
2711 /*
2712  * initialise the VMA slab
2713  */
2714 void __init mmap_init(void)
2715 {
2716         int ret;
2717
2718         ret = percpu_counter_init(&vm_committed_as, 0);
2719         VM_BUG_ON(ret);
2720 }