]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - mm/mmap.c
7d416055f08cf3605e2ae6a8160dc6b438a02c71
[~shefty/rdma-dev.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 #include <linux/uprobes.h>
34
35 #include <asm/uaccess.h>
36 #include <asm/cacheflush.h>
37 #include <asm/tlb.h>
38 #include <asm/mmu_context.h>
39
40 #include "internal.h"
41
42 #ifndef arch_mmap_check
43 #define arch_mmap_check(addr, len, flags)       (0)
44 #endif
45
46 #ifndef arch_rebalance_pgtables
47 #define arch_rebalance_pgtables(addr, len)              (addr)
48 #endif
49
50 static void unmap_region(struct mm_struct *mm,
51                 struct vm_area_struct *vma, struct vm_area_struct *prev,
52                 unsigned long start, unsigned long end);
53
54 /* description of effects of mapping type and prot in current implementation.
55  * this is due to the limited x86 page protection hardware.  The expected
56  * behavior is in parens:
57  *
58  * map_type     prot
59  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
60  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
61  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
62  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
63  *              
64  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
65  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
66  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
67  *
68  */
69 pgprot_t protection_map[16] = {
70         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
71         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
72 };
73
74 pgprot_t vm_get_page_prot(unsigned long vm_flags)
75 {
76         return __pgprot(pgprot_val(protection_map[vm_flags &
77                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
78                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
79 }
80 EXPORT_SYMBOL(vm_get_page_prot);
81
82 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
83 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
84 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
85 /*
86  * Make sure vm_committed_as in one cacheline and not cacheline shared with
87  * other variables. It can be updated by several CPUs frequently.
88  */
89 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
90
91 /*
92  * The global memory commitment made in the system can be a metric
93  * that can be used to drive ballooning decisions when Linux is hosted
94  * as a guest. On Hyper-V, the host implements a policy engine for dynamically
95  * balancing memory across competing virtual machines that are hosted.
96  * Several metrics drive this policy engine including the guest reported
97  * memory commitment.
98  */
99 unsigned long vm_memory_committed(void)
100 {
101         return percpu_counter_read_positive(&vm_committed_as);
102 }
103 EXPORT_SYMBOL_GPL(vm_memory_committed);
104
105 /*
106  * Check that a process has enough memory to allocate a new virtual
107  * mapping. 0 means there is enough memory for the allocation to
108  * succeed and -ENOMEM implies there is not.
109  *
110  * We currently support three overcommit policies, which are set via the
111  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
112  *
113  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
114  * Additional code 2002 Jul 20 by Robert Love.
115  *
116  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
117  *
118  * Note this is a helper function intended to be used by LSMs which
119  * wish to use this logic.
120  */
121 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
122 {
123         unsigned long free, allowed;
124
125         vm_acct_memory(pages);
126
127         /*
128          * Sometimes we want to use more memory than we have
129          */
130         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
131                 return 0;
132
133         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
134                 free = global_page_state(NR_FREE_PAGES);
135                 free += global_page_state(NR_FILE_PAGES);
136
137                 /*
138                  * shmem pages shouldn't be counted as free in this
139                  * case, they can't be purged, only swapped out, and
140                  * that won't affect the overall amount of available
141                  * memory in the system.
142                  */
143                 free -= global_page_state(NR_SHMEM);
144
145                 free += nr_swap_pages;
146
147                 /*
148                  * Any slabs which are created with the
149                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
150                  * which are reclaimable, under pressure.  The dentry
151                  * cache and most inode caches should fall into this
152                  */
153                 free += global_page_state(NR_SLAB_RECLAIMABLE);
154
155                 /*
156                  * Leave reserved pages. The pages are not for anonymous pages.
157                  */
158                 if (free <= totalreserve_pages)
159                         goto error;
160                 else
161                         free -= totalreserve_pages;
162
163                 /*
164                  * Leave the last 3% for root
165                  */
166                 if (!cap_sys_admin)
167                         free -= free / 32;
168
169                 if (free > pages)
170                         return 0;
171
172                 goto error;
173         }
174
175         allowed = (totalram_pages - hugetlb_total_pages())
176                 * sysctl_overcommit_ratio / 100;
177         /*
178          * Leave the last 3% for root
179          */
180         if (!cap_sys_admin)
181                 allowed -= allowed / 32;
182         allowed += total_swap_pages;
183
184         /* Don't let a single process grow too big:
185            leave 3% of the size of this process for other processes */
186         if (mm)
187                 allowed -= mm->total_vm / 32;
188
189         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
190                 return 0;
191 error:
192         vm_unacct_memory(pages);
193
194         return -ENOMEM;
195 }
196
197 /*
198  * Requires inode->i_mapping->i_mmap_mutex
199  */
200 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
201                 struct file *file, struct address_space *mapping)
202 {
203         if (vma->vm_flags & VM_DENYWRITE)
204                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
205         if (vma->vm_flags & VM_SHARED)
206                 mapping->i_mmap_writable--;
207
208         flush_dcache_mmap_lock(mapping);
209         if (unlikely(vma->vm_flags & VM_NONLINEAR))
210                 list_del_init(&vma->shared.nonlinear);
211         else
212                 vma_interval_tree_remove(vma, &mapping->i_mmap);
213         flush_dcache_mmap_unlock(mapping);
214 }
215
216 /*
217  * Unlink a file-based vm structure from its interval tree, to hide
218  * vma from rmap and vmtruncate before freeing its page tables.
219  */
220 void unlink_file_vma(struct vm_area_struct *vma)
221 {
222         struct file *file = vma->vm_file;
223
224         if (file) {
225                 struct address_space *mapping = file->f_mapping;
226                 mutex_lock(&mapping->i_mmap_mutex);
227                 __remove_shared_vm_struct(vma, file, mapping);
228                 mutex_unlock(&mapping->i_mmap_mutex);
229         }
230 }
231
232 /*
233  * Close a vm structure and free it, returning the next.
234  */
235 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
236 {
237         struct vm_area_struct *next = vma->vm_next;
238
239         might_sleep();
240         if (vma->vm_ops && vma->vm_ops->close)
241                 vma->vm_ops->close(vma);
242         if (vma->vm_file)
243                 fput(vma->vm_file);
244         mpol_put(vma_policy(vma));
245         kmem_cache_free(vm_area_cachep, vma);
246         return next;
247 }
248
249 static unsigned long do_brk(unsigned long addr, unsigned long len);
250
251 SYSCALL_DEFINE1(brk, unsigned long, brk)
252 {
253         unsigned long rlim, retval;
254         unsigned long newbrk, oldbrk;
255         struct mm_struct *mm = current->mm;
256         unsigned long min_brk;
257
258         down_write(&mm->mmap_sem);
259
260 #ifdef CONFIG_COMPAT_BRK
261         /*
262          * CONFIG_COMPAT_BRK can still be overridden by setting
263          * randomize_va_space to 2, which will still cause mm->start_brk
264          * to be arbitrarily shifted
265          */
266         if (current->brk_randomized)
267                 min_brk = mm->start_brk;
268         else
269                 min_brk = mm->end_data;
270 #else
271         min_brk = mm->start_brk;
272 #endif
273         if (brk < min_brk)
274                 goto out;
275
276         /*
277          * Check against rlimit here. If this check is done later after the test
278          * of oldbrk with newbrk then it can escape the test and let the data
279          * segment grow beyond its set limit the in case where the limit is
280          * not page aligned -Ram Gupta
281          */
282         rlim = rlimit(RLIMIT_DATA);
283         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
284                         (mm->end_data - mm->start_data) > rlim)
285                 goto out;
286
287         newbrk = PAGE_ALIGN(brk);
288         oldbrk = PAGE_ALIGN(mm->brk);
289         if (oldbrk == newbrk)
290                 goto set_brk;
291
292         /* Always allow shrinking brk. */
293         if (brk <= mm->brk) {
294                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
295                         goto set_brk;
296                 goto out;
297         }
298
299         /* Check against existing mmap mappings. */
300         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
301                 goto out;
302
303         /* Ok, looks good - let it rip. */
304         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
305                 goto out;
306 set_brk:
307         mm->brk = brk;
308 out:
309         retval = mm->brk;
310         up_write(&mm->mmap_sem);
311         return retval;
312 }
313
314 #ifdef CONFIG_DEBUG_VM_RB
315 static int browse_rb(struct rb_root *root)
316 {
317         int i = 0, j;
318         struct rb_node *nd, *pn = NULL;
319         unsigned long prev = 0, pend = 0;
320
321         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
322                 struct vm_area_struct *vma;
323                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
324                 if (vma->vm_start < prev)
325                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
326                 if (vma->vm_start < pend)
327                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
328                 if (vma->vm_start > vma->vm_end)
329                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
330                 i++;
331                 pn = nd;
332                 prev = vma->vm_start;
333                 pend = vma->vm_end;
334         }
335         j = 0;
336         for (nd = pn; nd; nd = rb_prev(nd)) {
337                 j++;
338         }
339         if (i != j)
340                 printk("backwards %d, forwards %d\n", j, i), i = 0;
341         return i;
342 }
343
344 void validate_mm(struct mm_struct *mm)
345 {
346         int bug = 0;
347         int i = 0;
348         struct vm_area_struct *vma = mm->mmap;
349         while (vma) {
350                 struct anon_vma_chain *avc;
351                 vma_lock_anon_vma(vma);
352                 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
353                         anon_vma_interval_tree_verify(avc);
354                 vma_unlock_anon_vma(vma);
355                 vma = vma->vm_next;
356                 i++;
357         }
358         if (i != mm->map_count)
359                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
360         i = browse_rb(&mm->mm_rb);
361         if (i != mm->map_count)
362                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
363         BUG_ON(bug);
364 }
365 #else
366 #define validate_mm(mm) do { } while (0)
367 #endif
368
369 /*
370  * vma has some anon_vma assigned, and is already inserted on that
371  * anon_vma's interval trees.
372  *
373  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
374  * vma must be removed from the anon_vma's interval trees using
375  * anon_vma_interval_tree_pre_update_vma().
376  *
377  * After the update, the vma will be reinserted using
378  * anon_vma_interval_tree_post_update_vma().
379  *
380  * The entire update must be protected by exclusive mmap_sem and by
381  * the root anon_vma's mutex.
382  */
383 static inline void
384 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
385 {
386         struct anon_vma_chain *avc;
387
388         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
389                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
390 }
391
392 static inline void
393 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
394 {
395         struct anon_vma_chain *avc;
396
397         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
398                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
399 }
400
401 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
402                 unsigned long end, struct vm_area_struct **pprev,
403                 struct rb_node ***rb_link, struct rb_node **rb_parent)
404 {
405         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
406
407         __rb_link = &mm->mm_rb.rb_node;
408         rb_prev = __rb_parent = NULL;
409
410         while (*__rb_link) {
411                 struct vm_area_struct *vma_tmp;
412
413                 __rb_parent = *__rb_link;
414                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
415
416                 if (vma_tmp->vm_end > addr) {
417                         /* Fail if an existing vma overlaps the area */
418                         if (vma_tmp->vm_start < end)
419                                 return -ENOMEM;
420                         __rb_link = &__rb_parent->rb_left;
421                 } else {
422                         rb_prev = __rb_parent;
423                         __rb_link = &__rb_parent->rb_right;
424                 }
425         }
426
427         *pprev = NULL;
428         if (rb_prev)
429                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
430         *rb_link = __rb_link;
431         *rb_parent = __rb_parent;
432         return 0;
433 }
434
435 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
436                 struct rb_node **rb_link, struct rb_node *rb_parent)
437 {
438         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
439         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
440 }
441
442 static void __vma_link_file(struct vm_area_struct *vma)
443 {
444         struct file *file;
445
446         file = vma->vm_file;
447         if (file) {
448                 struct address_space *mapping = file->f_mapping;
449
450                 if (vma->vm_flags & VM_DENYWRITE)
451                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
452                 if (vma->vm_flags & VM_SHARED)
453                         mapping->i_mmap_writable++;
454
455                 flush_dcache_mmap_lock(mapping);
456                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
457                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
458                 else
459                         vma_interval_tree_insert(vma, &mapping->i_mmap);
460                 flush_dcache_mmap_unlock(mapping);
461         }
462 }
463
464 static void
465 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
466         struct vm_area_struct *prev, struct rb_node **rb_link,
467         struct rb_node *rb_parent)
468 {
469         __vma_link_list(mm, vma, prev, rb_parent);
470         __vma_link_rb(mm, vma, rb_link, rb_parent);
471 }
472
473 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
474                         struct vm_area_struct *prev, struct rb_node **rb_link,
475                         struct rb_node *rb_parent)
476 {
477         struct address_space *mapping = NULL;
478
479         if (vma->vm_file)
480                 mapping = vma->vm_file->f_mapping;
481
482         if (mapping)
483                 mutex_lock(&mapping->i_mmap_mutex);
484
485         __vma_link(mm, vma, prev, rb_link, rb_parent);
486         __vma_link_file(vma);
487
488         if (mapping)
489                 mutex_unlock(&mapping->i_mmap_mutex);
490
491         mm->map_count++;
492         validate_mm(mm);
493 }
494
495 /*
496  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
497  * mm's list and rbtree.  It has already been inserted into the interval tree.
498  */
499 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
500 {
501         struct vm_area_struct *prev;
502         struct rb_node **rb_link, *rb_parent;
503
504         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
505                            &prev, &rb_link, &rb_parent))
506                 BUG();
507         __vma_link(mm, vma, prev, rb_link, rb_parent);
508         mm->map_count++;
509 }
510
511 static inline void
512 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
513                 struct vm_area_struct *prev)
514 {
515         struct vm_area_struct *next = vma->vm_next;
516
517         prev->vm_next = next;
518         if (next)
519                 next->vm_prev = prev;
520         rb_erase(&vma->vm_rb, &mm->mm_rb);
521         if (mm->mmap_cache == vma)
522                 mm->mmap_cache = prev;
523 }
524
525 /*
526  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
527  * is already present in an i_mmap tree without adjusting the tree.
528  * The following helper function should be used when such adjustments
529  * are necessary.  The "insert" vma (if any) is to be inserted
530  * before we drop the necessary locks.
531  */
532 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
533         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
534 {
535         struct mm_struct *mm = vma->vm_mm;
536         struct vm_area_struct *next = vma->vm_next;
537         struct vm_area_struct *importer = NULL;
538         struct address_space *mapping = NULL;
539         struct rb_root *root = NULL;
540         struct anon_vma *anon_vma = NULL;
541         struct file *file = vma->vm_file;
542         long adjust_next = 0;
543         int remove_next = 0;
544
545         if (next && !insert) {
546                 struct vm_area_struct *exporter = NULL;
547
548                 if (end >= next->vm_end) {
549                         /*
550                          * vma expands, overlapping all the next, and
551                          * perhaps the one after too (mprotect case 6).
552                          */
553 again:                  remove_next = 1 + (end > next->vm_end);
554                         end = next->vm_end;
555                         exporter = next;
556                         importer = vma;
557                 } else if (end > next->vm_start) {
558                         /*
559                          * vma expands, overlapping part of the next:
560                          * mprotect case 5 shifting the boundary up.
561                          */
562                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
563                         exporter = next;
564                         importer = vma;
565                 } else if (end < vma->vm_end) {
566                         /*
567                          * vma shrinks, and !insert tells it's not
568                          * split_vma inserting another: so it must be
569                          * mprotect case 4 shifting the boundary down.
570                          */
571                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
572                         exporter = vma;
573                         importer = next;
574                 }
575
576                 /*
577                  * Easily overlooked: when mprotect shifts the boundary,
578                  * make sure the expanding vma has anon_vma set if the
579                  * shrinking vma had, to cover any anon pages imported.
580                  */
581                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
582                         if (anon_vma_clone(importer, exporter))
583                                 return -ENOMEM;
584                         importer->anon_vma = exporter->anon_vma;
585                 }
586         }
587
588         if (file) {
589                 mapping = file->f_mapping;
590                 if (!(vma->vm_flags & VM_NONLINEAR)) {
591                         root = &mapping->i_mmap;
592                         uprobe_munmap(vma, vma->vm_start, vma->vm_end);
593
594                         if (adjust_next)
595                                 uprobe_munmap(next, next->vm_start,
596                                                         next->vm_end);
597                 }
598
599                 mutex_lock(&mapping->i_mmap_mutex);
600                 if (insert) {
601                         /*
602                          * Put into interval tree now, so instantiated pages
603                          * are visible to arm/parisc __flush_dcache_page
604                          * throughout; but we cannot insert into address
605                          * space until vma start or end is updated.
606                          */
607                         __vma_link_file(insert);
608                 }
609         }
610
611         vma_adjust_trans_huge(vma, start, end, adjust_next);
612
613         anon_vma = vma->anon_vma;
614         if (!anon_vma && adjust_next)
615                 anon_vma = next->anon_vma;
616         if (anon_vma) {
617                 VM_BUG_ON(adjust_next && next->anon_vma &&
618                           anon_vma != next->anon_vma);
619                 anon_vma_lock(anon_vma);
620                 anon_vma_interval_tree_pre_update_vma(vma);
621                 if (adjust_next)
622                         anon_vma_interval_tree_pre_update_vma(next);
623         }
624
625         if (root) {
626                 flush_dcache_mmap_lock(mapping);
627                 vma_interval_tree_remove(vma, root);
628                 if (adjust_next)
629                         vma_interval_tree_remove(next, root);
630         }
631
632         vma->vm_start = start;
633         vma->vm_end = end;
634         vma->vm_pgoff = pgoff;
635         if (adjust_next) {
636                 next->vm_start += adjust_next << PAGE_SHIFT;
637                 next->vm_pgoff += adjust_next;
638         }
639
640         if (root) {
641                 if (adjust_next)
642                         vma_interval_tree_insert(next, root);
643                 vma_interval_tree_insert(vma, root);
644                 flush_dcache_mmap_unlock(mapping);
645         }
646
647         if (remove_next) {
648                 /*
649                  * vma_merge has merged next into vma, and needs
650                  * us to remove next before dropping the locks.
651                  */
652                 __vma_unlink(mm, next, vma);
653                 if (file)
654                         __remove_shared_vm_struct(next, file, mapping);
655         } else if (insert) {
656                 /*
657                  * split_vma has split insert from vma, and needs
658                  * us to insert it before dropping the locks
659                  * (it may either follow vma or precede it).
660                  */
661                 __insert_vm_struct(mm, insert);
662         }
663
664         if (anon_vma) {
665                 anon_vma_interval_tree_post_update_vma(vma);
666                 if (adjust_next)
667                         anon_vma_interval_tree_post_update_vma(next);
668                 anon_vma_unlock(anon_vma);
669         }
670         if (mapping)
671                 mutex_unlock(&mapping->i_mmap_mutex);
672
673         if (root) {
674                 uprobe_mmap(vma);
675
676                 if (adjust_next)
677                         uprobe_mmap(next);
678         }
679
680         if (remove_next) {
681                 if (file) {
682                         uprobe_munmap(next, next->vm_start, next->vm_end);
683                         fput(file);
684                 }
685                 if (next->anon_vma)
686                         anon_vma_merge(vma, next);
687                 mm->map_count--;
688                 mpol_put(vma_policy(next));
689                 kmem_cache_free(vm_area_cachep, next);
690                 /*
691                  * In mprotect's case 6 (see comments on vma_merge),
692                  * we must remove another next too. It would clutter
693                  * up the code too much to do both in one go.
694                  */
695                 if (remove_next == 2) {
696                         next = vma->vm_next;
697                         goto again;
698                 }
699         }
700         if (insert && file)
701                 uprobe_mmap(insert);
702
703         validate_mm(mm);
704
705         return 0;
706 }
707
708 /*
709  * If the vma has a ->close operation then the driver probably needs to release
710  * per-vma resources, so we don't attempt to merge those.
711  */
712 static inline int is_mergeable_vma(struct vm_area_struct *vma,
713                         struct file *file, unsigned long vm_flags)
714 {
715         if (vma->vm_flags ^ vm_flags)
716                 return 0;
717         if (vma->vm_file != file)
718                 return 0;
719         if (vma->vm_ops && vma->vm_ops->close)
720                 return 0;
721         return 1;
722 }
723
724 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
725                                         struct anon_vma *anon_vma2,
726                                         struct vm_area_struct *vma)
727 {
728         /*
729          * The list_is_singular() test is to avoid merging VMA cloned from
730          * parents. This can improve scalability caused by anon_vma lock.
731          */
732         if ((!anon_vma1 || !anon_vma2) && (!vma ||
733                 list_is_singular(&vma->anon_vma_chain)))
734                 return 1;
735         return anon_vma1 == anon_vma2;
736 }
737
738 /*
739  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
740  * in front of (at a lower virtual address and file offset than) the vma.
741  *
742  * We cannot merge two vmas if they have differently assigned (non-NULL)
743  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
744  *
745  * We don't check here for the merged mmap wrapping around the end of pagecache
746  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
747  * wrap, nor mmaps which cover the final page at index -1UL.
748  */
749 static int
750 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
751         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
752 {
753         if (is_mergeable_vma(vma, file, vm_flags) &&
754             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
755                 if (vma->vm_pgoff == vm_pgoff)
756                         return 1;
757         }
758         return 0;
759 }
760
761 /*
762  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
763  * beyond (at a higher virtual address and file offset than) the vma.
764  *
765  * We cannot merge two vmas if they have differently assigned (non-NULL)
766  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
767  */
768 static int
769 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
770         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
771 {
772         if (is_mergeable_vma(vma, file, vm_flags) &&
773             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
774                 pgoff_t vm_pglen;
775                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
776                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
777                         return 1;
778         }
779         return 0;
780 }
781
782 /*
783  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
784  * whether that can be merged with its predecessor or its successor.
785  * Or both (it neatly fills a hole).
786  *
787  * In most cases - when called for mmap, brk or mremap - [addr,end) is
788  * certain not to be mapped by the time vma_merge is called; but when
789  * called for mprotect, it is certain to be already mapped (either at
790  * an offset within prev, or at the start of next), and the flags of
791  * this area are about to be changed to vm_flags - and the no-change
792  * case has already been eliminated.
793  *
794  * The following mprotect cases have to be considered, where AAAA is
795  * the area passed down from mprotect_fixup, never extending beyond one
796  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
797  *
798  *     AAAA             AAAA                AAAA          AAAA
799  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
800  *    cannot merge    might become    might become    might become
801  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
802  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
803  *    mremap move:                                    PPPPNNNNNNNN 8
804  *        AAAA
805  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
806  *    might become    case 1 below    case 2 below    case 3 below
807  *
808  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
809  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
810  */
811 struct vm_area_struct *vma_merge(struct mm_struct *mm,
812                         struct vm_area_struct *prev, unsigned long addr,
813                         unsigned long end, unsigned long vm_flags,
814                         struct anon_vma *anon_vma, struct file *file,
815                         pgoff_t pgoff, struct mempolicy *policy)
816 {
817         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
818         struct vm_area_struct *area, *next;
819         int err;
820
821         /*
822          * We later require that vma->vm_flags == vm_flags,
823          * so this tests vma->vm_flags & VM_SPECIAL, too.
824          */
825         if (vm_flags & VM_SPECIAL)
826                 return NULL;
827
828         if (prev)
829                 next = prev->vm_next;
830         else
831                 next = mm->mmap;
832         area = next;
833         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
834                 next = next->vm_next;
835
836         /*
837          * Can it merge with the predecessor?
838          */
839         if (prev && prev->vm_end == addr &&
840                         mpol_equal(vma_policy(prev), policy) &&
841                         can_vma_merge_after(prev, vm_flags,
842                                                 anon_vma, file, pgoff)) {
843                 /*
844                  * OK, it can.  Can we now merge in the successor as well?
845                  */
846                 if (next && end == next->vm_start &&
847                                 mpol_equal(policy, vma_policy(next)) &&
848                                 can_vma_merge_before(next, vm_flags,
849                                         anon_vma, file, pgoff+pglen) &&
850                                 is_mergeable_anon_vma(prev->anon_vma,
851                                                       next->anon_vma, NULL)) {
852                                                         /* cases 1, 6 */
853                         err = vma_adjust(prev, prev->vm_start,
854                                 next->vm_end, prev->vm_pgoff, NULL);
855                 } else                                  /* cases 2, 5, 7 */
856                         err = vma_adjust(prev, prev->vm_start,
857                                 end, prev->vm_pgoff, NULL);
858                 if (err)
859                         return NULL;
860                 khugepaged_enter_vma_merge(prev);
861                 return prev;
862         }
863
864         /*
865          * Can this new request be merged in front of next?
866          */
867         if (next && end == next->vm_start &&
868                         mpol_equal(policy, vma_policy(next)) &&
869                         can_vma_merge_before(next, vm_flags,
870                                         anon_vma, file, pgoff+pglen)) {
871                 if (prev && addr < prev->vm_end)        /* case 4 */
872                         err = vma_adjust(prev, prev->vm_start,
873                                 addr, prev->vm_pgoff, NULL);
874                 else                                    /* cases 3, 8 */
875                         err = vma_adjust(area, addr, next->vm_end,
876                                 next->vm_pgoff - pglen, NULL);
877                 if (err)
878                         return NULL;
879                 khugepaged_enter_vma_merge(area);
880                 return area;
881         }
882
883         return NULL;
884 }
885
886 /*
887  * Rough compatbility check to quickly see if it's even worth looking
888  * at sharing an anon_vma.
889  *
890  * They need to have the same vm_file, and the flags can only differ
891  * in things that mprotect may change.
892  *
893  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
894  * we can merge the two vma's. For example, we refuse to merge a vma if
895  * there is a vm_ops->close() function, because that indicates that the
896  * driver is doing some kind of reference counting. But that doesn't
897  * really matter for the anon_vma sharing case.
898  */
899 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
900 {
901         return a->vm_end == b->vm_start &&
902                 mpol_equal(vma_policy(a), vma_policy(b)) &&
903                 a->vm_file == b->vm_file &&
904                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
905                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
906 }
907
908 /*
909  * Do some basic sanity checking to see if we can re-use the anon_vma
910  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
911  * the same as 'old', the other will be the new one that is trying
912  * to share the anon_vma.
913  *
914  * NOTE! This runs with mm_sem held for reading, so it is possible that
915  * the anon_vma of 'old' is concurrently in the process of being set up
916  * by another page fault trying to merge _that_. But that's ok: if it
917  * is being set up, that automatically means that it will be a singleton
918  * acceptable for merging, so we can do all of this optimistically. But
919  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
920  *
921  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
922  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
923  * is to return an anon_vma that is "complex" due to having gone through
924  * a fork).
925  *
926  * We also make sure that the two vma's are compatible (adjacent,
927  * and with the same memory policies). That's all stable, even with just
928  * a read lock on the mm_sem.
929  */
930 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
931 {
932         if (anon_vma_compatible(a, b)) {
933                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
934
935                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
936                         return anon_vma;
937         }
938         return NULL;
939 }
940
941 /*
942  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
943  * neighbouring vmas for a suitable anon_vma, before it goes off
944  * to allocate a new anon_vma.  It checks because a repetitive
945  * sequence of mprotects and faults may otherwise lead to distinct
946  * anon_vmas being allocated, preventing vma merge in subsequent
947  * mprotect.
948  */
949 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
950 {
951         struct anon_vma *anon_vma;
952         struct vm_area_struct *near;
953
954         near = vma->vm_next;
955         if (!near)
956                 goto try_prev;
957
958         anon_vma = reusable_anon_vma(near, vma, near);
959         if (anon_vma)
960                 return anon_vma;
961 try_prev:
962         near = vma->vm_prev;
963         if (!near)
964                 goto none;
965
966         anon_vma = reusable_anon_vma(near, near, vma);
967         if (anon_vma)
968                 return anon_vma;
969 none:
970         /*
971          * There's no absolute need to look only at touching neighbours:
972          * we could search further afield for "compatible" anon_vmas.
973          * But it would probably just be a waste of time searching,
974          * or lead to too many vmas hanging off the same anon_vma.
975          * We're trying to allow mprotect remerging later on,
976          * not trying to minimize memory used for anon_vmas.
977          */
978         return NULL;
979 }
980
981 #ifdef CONFIG_PROC_FS
982 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
983                                                 struct file *file, long pages)
984 {
985         const unsigned long stack_flags
986                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
987
988         mm->total_vm += pages;
989
990         if (file) {
991                 mm->shared_vm += pages;
992                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
993                         mm->exec_vm += pages;
994         } else if (flags & stack_flags)
995                 mm->stack_vm += pages;
996 }
997 #endif /* CONFIG_PROC_FS */
998
999 /*
1000  * If a hint addr is less than mmap_min_addr change hint to be as
1001  * low as possible but still greater than mmap_min_addr
1002  */
1003 static inline unsigned long round_hint_to_min(unsigned long hint)
1004 {
1005         hint &= PAGE_MASK;
1006         if (((void *)hint != NULL) &&
1007             (hint < mmap_min_addr))
1008                 return PAGE_ALIGN(mmap_min_addr);
1009         return hint;
1010 }
1011
1012 /*
1013  * The caller must hold down_write(&current->mm->mmap_sem).
1014  */
1015
1016 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1017                         unsigned long len, unsigned long prot,
1018                         unsigned long flags, unsigned long pgoff)
1019 {
1020         struct mm_struct * mm = current->mm;
1021         struct inode *inode;
1022         vm_flags_t vm_flags;
1023
1024         /*
1025          * Does the application expect PROT_READ to imply PROT_EXEC?
1026          *
1027          * (the exception is when the underlying filesystem is noexec
1028          *  mounted, in which case we dont add PROT_EXEC.)
1029          */
1030         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1031                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1032                         prot |= PROT_EXEC;
1033
1034         if (!len)
1035                 return -EINVAL;
1036
1037         if (!(flags & MAP_FIXED))
1038                 addr = round_hint_to_min(addr);
1039
1040         /* Careful about overflows.. */
1041         len = PAGE_ALIGN(len);
1042         if (!len)
1043                 return -ENOMEM;
1044
1045         /* offset overflow? */
1046         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1047                return -EOVERFLOW;
1048
1049         /* Too many mappings? */
1050         if (mm->map_count > sysctl_max_map_count)
1051                 return -ENOMEM;
1052
1053         /* Obtain the address to map to. we verify (or select) it and ensure
1054          * that it represents a valid section of the address space.
1055          */
1056         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1057         if (addr & ~PAGE_MASK)
1058                 return addr;
1059
1060         /* Do simple checking here so the lower-level routines won't have
1061          * to. we assume access permissions have been handled by the open
1062          * of the memory object, so we don't do any here.
1063          */
1064         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1065                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1066
1067         if (flags & MAP_LOCKED)
1068                 if (!can_do_mlock())
1069                         return -EPERM;
1070
1071         /* mlock MCL_FUTURE? */
1072         if (vm_flags & VM_LOCKED) {
1073                 unsigned long locked, lock_limit;
1074                 locked = len >> PAGE_SHIFT;
1075                 locked += mm->locked_vm;
1076                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1077                 lock_limit >>= PAGE_SHIFT;
1078                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1079                         return -EAGAIN;
1080         }
1081
1082         inode = file ? file->f_path.dentry->d_inode : NULL;
1083
1084         if (file) {
1085                 switch (flags & MAP_TYPE) {
1086                 case MAP_SHARED:
1087                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1088                                 return -EACCES;
1089
1090                         /*
1091                          * Make sure we don't allow writing to an append-only
1092                          * file..
1093                          */
1094                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1095                                 return -EACCES;
1096
1097                         /*
1098                          * Make sure there are no mandatory locks on the file.
1099                          */
1100                         if (locks_verify_locked(inode))
1101                                 return -EAGAIN;
1102
1103                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1104                         if (!(file->f_mode & FMODE_WRITE))
1105                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1106
1107                         /* fall through */
1108                 case MAP_PRIVATE:
1109                         if (!(file->f_mode & FMODE_READ))
1110                                 return -EACCES;
1111                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1112                                 if (vm_flags & VM_EXEC)
1113                                         return -EPERM;
1114                                 vm_flags &= ~VM_MAYEXEC;
1115                         }
1116
1117                         if (!file->f_op || !file->f_op->mmap)
1118                                 return -ENODEV;
1119                         break;
1120
1121                 default:
1122                         return -EINVAL;
1123                 }
1124         } else {
1125                 switch (flags & MAP_TYPE) {
1126                 case MAP_SHARED:
1127                         /*
1128                          * Ignore pgoff.
1129                          */
1130                         pgoff = 0;
1131                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1132                         break;
1133                 case MAP_PRIVATE:
1134                         /*
1135                          * Set pgoff according to addr for anon_vma.
1136                          */
1137                         pgoff = addr >> PAGE_SHIFT;
1138                         break;
1139                 default:
1140                         return -EINVAL;
1141                 }
1142         }
1143
1144         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1145 }
1146
1147 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1148                 unsigned long, prot, unsigned long, flags,
1149                 unsigned long, fd, unsigned long, pgoff)
1150 {
1151         struct file *file = NULL;
1152         unsigned long retval = -EBADF;
1153
1154         if (!(flags & MAP_ANONYMOUS)) {
1155                 audit_mmap_fd(fd, flags);
1156                 if (unlikely(flags & MAP_HUGETLB))
1157                         return -EINVAL;
1158                 file = fget(fd);
1159                 if (!file)
1160                         goto out;
1161         } else if (flags & MAP_HUGETLB) {
1162                 struct user_struct *user = NULL;
1163                 /*
1164                  * VM_NORESERVE is used because the reservations will be
1165                  * taken when vm_ops->mmap() is called
1166                  * A dummy user value is used because we are not locking
1167                  * memory so no accounting is necessary
1168                  */
1169                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1170                                                 VM_NORESERVE, &user,
1171                                                 HUGETLB_ANONHUGE_INODE);
1172                 if (IS_ERR(file))
1173                         return PTR_ERR(file);
1174         }
1175
1176         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1177
1178         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1179         if (file)
1180                 fput(file);
1181 out:
1182         return retval;
1183 }
1184
1185 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1186 struct mmap_arg_struct {
1187         unsigned long addr;
1188         unsigned long len;
1189         unsigned long prot;
1190         unsigned long flags;
1191         unsigned long fd;
1192         unsigned long offset;
1193 };
1194
1195 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1196 {
1197         struct mmap_arg_struct a;
1198
1199         if (copy_from_user(&a, arg, sizeof(a)))
1200                 return -EFAULT;
1201         if (a.offset & ~PAGE_MASK)
1202                 return -EINVAL;
1203
1204         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1205                               a.offset >> PAGE_SHIFT);
1206 }
1207 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1208
1209 /*
1210  * Some shared mappigns will want the pages marked read-only
1211  * to track write events. If so, we'll downgrade vm_page_prot
1212  * to the private version (using protection_map[] without the
1213  * VM_SHARED bit).
1214  */
1215 int vma_wants_writenotify(struct vm_area_struct *vma)
1216 {
1217         vm_flags_t vm_flags = vma->vm_flags;
1218
1219         /* If it was private or non-writable, the write bit is already clear */
1220         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1221                 return 0;
1222
1223         /* The backer wishes to know when pages are first written to? */
1224         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1225                 return 1;
1226
1227         /* The open routine did something to the protections already? */
1228         if (pgprot_val(vma->vm_page_prot) !=
1229             pgprot_val(vm_get_page_prot(vm_flags)))
1230                 return 0;
1231
1232         /* Specialty mapping? */
1233         if (vm_flags & VM_PFNMAP)
1234                 return 0;
1235
1236         /* Can the mapping track the dirty pages? */
1237         return vma->vm_file && vma->vm_file->f_mapping &&
1238                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1239 }
1240
1241 /*
1242  * We account for memory if it's a private writeable mapping,
1243  * not hugepages and VM_NORESERVE wasn't set.
1244  */
1245 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1246 {
1247         /*
1248          * hugetlb has its own accounting separate from the core VM
1249          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1250          */
1251         if (file && is_file_hugepages(file))
1252                 return 0;
1253
1254         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1255 }
1256
1257 unsigned long mmap_region(struct file *file, unsigned long addr,
1258                           unsigned long len, unsigned long flags,
1259                           vm_flags_t vm_flags, unsigned long pgoff)
1260 {
1261         struct mm_struct *mm = current->mm;
1262         struct vm_area_struct *vma, *prev;
1263         int correct_wcount = 0;
1264         int error;
1265         struct rb_node **rb_link, *rb_parent;
1266         unsigned long charged = 0;
1267         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1268
1269         /* Clear old maps */
1270         error = -ENOMEM;
1271 munmap_back:
1272         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
1273                 if (do_munmap(mm, addr, len))
1274                         return -ENOMEM;
1275                 goto munmap_back;
1276         }
1277
1278         /* Check against address space limit. */
1279         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1280                 return -ENOMEM;
1281
1282         /*
1283          * Set 'VM_NORESERVE' if we should not account for the
1284          * memory use of this mapping.
1285          */
1286         if ((flags & MAP_NORESERVE)) {
1287                 /* We honor MAP_NORESERVE if allowed to overcommit */
1288                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1289                         vm_flags |= VM_NORESERVE;
1290
1291                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1292                 if (file && is_file_hugepages(file))
1293                         vm_flags |= VM_NORESERVE;
1294         }
1295
1296         /*
1297          * Private writable mapping: check memory availability
1298          */
1299         if (accountable_mapping(file, vm_flags)) {
1300                 charged = len >> PAGE_SHIFT;
1301                 if (security_vm_enough_memory_mm(mm, charged))
1302                         return -ENOMEM;
1303                 vm_flags |= VM_ACCOUNT;
1304         }
1305
1306         /*
1307          * Can we just expand an old mapping?
1308          */
1309         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1310         if (vma)
1311                 goto out;
1312
1313         /*
1314          * Determine the object being mapped and call the appropriate
1315          * specific mapper. the address has already been validated, but
1316          * not unmapped, but the maps are removed from the list.
1317          */
1318         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1319         if (!vma) {
1320                 error = -ENOMEM;
1321                 goto unacct_error;
1322         }
1323
1324         vma->vm_mm = mm;
1325         vma->vm_start = addr;
1326         vma->vm_end = addr + len;
1327         vma->vm_flags = vm_flags;
1328         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1329         vma->vm_pgoff = pgoff;
1330         INIT_LIST_HEAD(&vma->anon_vma_chain);
1331
1332         error = -EINVAL;        /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1333
1334         if (file) {
1335                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1336                         goto free_vma;
1337                 if (vm_flags & VM_DENYWRITE) {
1338                         error = deny_write_access(file);
1339                         if (error)
1340                                 goto free_vma;
1341                         correct_wcount = 1;
1342                 }
1343                 vma->vm_file = get_file(file);
1344                 error = file->f_op->mmap(file, vma);
1345                 if (error)
1346                         goto unmap_and_free_vma;
1347
1348                 /* Can addr have changed??
1349                  *
1350                  * Answer: Yes, several device drivers can do it in their
1351                  *         f_op->mmap method. -DaveM
1352                  */
1353                 addr = vma->vm_start;
1354                 pgoff = vma->vm_pgoff;
1355                 vm_flags = vma->vm_flags;
1356         } else if (vm_flags & VM_SHARED) {
1357                 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1358                         goto free_vma;
1359                 error = shmem_zero_setup(vma);
1360                 if (error)
1361                         goto free_vma;
1362         }
1363
1364         if (vma_wants_writenotify(vma)) {
1365                 pgprot_t pprot = vma->vm_page_prot;
1366
1367                 /* Can vma->vm_page_prot have changed??
1368                  *
1369                  * Answer: Yes, drivers may have changed it in their
1370                  *         f_op->mmap method.
1371                  *
1372                  * Ensures that vmas marked as uncached stay that way.
1373                  */
1374                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1375                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1376                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1377         }
1378
1379         vma_link(mm, vma, prev, rb_link, rb_parent);
1380         file = vma->vm_file;
1381
1382         /* Once vma denies write, undo our temporary denial count */
1383         if (correct_wcount)
1384                 atomic_inc(&inode->i_writecount);
1385 out:
1386         perf_event_mmap(vma);
1387
1388         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1389         if (vm_flags & VM_LOCKED) {
1390                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1391                         mm->locked_vm += (len >> PAGE_SHIFT);
1392         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1393                 make_pages_present(addr, addr + len);
1394
1395         if (file)
1396                 uprobe_mmap(vma);
1397
1398         return addr;
1399
1400 unmap_and_free_vma:
1401         if (correct_wcount)
1402                 atomic_inc(&inode->i_writecount);
1403         vma->vm_file = NULL;
1404         fput(file);
1405
1406         /* Undo any partial mapping done by a device driver. */
1407         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1408         charged = 0;
1409 free_vma:
1410         kmem_cache_free(vm_area_cachep, vma);
1411 unacct_error:
1412         if (charged)
1413                 vm_unacct_memory(charged);
1414         return error;
1415 }
1416
1417 /* Get an address range which is currently unmapped.
1418  * For shmat() with addr=0.
1419  *
1420  * Ugly calling convention alert:
1421  * Return value with the low bits set means error value,
1422  * ie
1423  *      if (ret & ~PAGE_MASK)
1424  *              error = ret;
1425  *
1426  * This function "knows" that -ENOMEM has the bits set.
1427  */
1428 #ifndef HAVE_ARCH_UNMAPPED_AREA
1429 unsigned long
1430 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1431                 unsigned long len, unsigned long pgoff, unsigned long flags)
1432 {
1433         struct mm_struct *mm = current->mm;
1434         struct vm_area_struct *vma;
1435         unsigned long start_addr;
1436
1437         if (len > TASK_SIZE)
1438                 return -ENOMEM;
1439
1440         if (flags & MAP_FIXED)
1441                 return addr;
1442
1443         if (addr) {
1444                 addr = PAGE_ALIGN(addr);
1445                 vma = find_vma(mm, addr);
1446                 if (TASK_SIZE - len >= addr &&
1447                     (!vma || addr + len <= vma->vm_start))
1448                         return addr;
1449         }
1450         if (len > mm->cached_hole_size) {
1451                 start_addr = addr = mm->free_area_cache;
1452         } else {
1453                 start_addr = addr = TASK_UNMAPPED_BASE;
1454                 mm->cached_hole_size = 0;
1455         }
1456
1457 full_search:
1458         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1459                 /* At this point:  (!vma || addr < vma->vm_end). */
1460                 if (TASK_SIZE - len < addr) {
1461                         /*
1462                          * Start a new search - just in case we missed
1463                          * some holes.
1464                          */
1465                         if (start_addr != TASK_UNMAPPED_BASE) {
1466                                 addr = TASK_UNMAPPED_BASE;
1467                                 start_addr = addr;
1468                                 mm->cached_hole_size = 0;
1469                                 goto full_search;
1470                         }
1471                         return -ENOMEM;
1472                 }
1473                 if (!vma || addr + len <= vma->vm_start) {
1474                         /*
1475                          * Remember the place where we stopped the search:
1476                          */
1477                         mm->free_area_cache = addr + len;
1478                         return addr;
1479                 }
1480                 if (addr + mm->cached_hole_size < vma->vm_start)
1481                         mm->cached_hole_size = vma->vm_start - addr;
1482                 addr = vma->vm_end;
1483         }
1484 }
1485 #endif  
1486
1487 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1488 {
1489         /*
1490          * Is this a new hole at the lowest possible address?
1491          */
1492         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1493                 mm->free_area_cache = addr;
1494 }
1495
1496 /*
1497  * This mmap-allocator allocates new areas top-down from below the
1498  * stack's low limit (the base):
1499  */
1500 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1501 unsigned long
1502 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1503                           const unsigned long len, const unsigned long pgoff,
1504                           const unsigned long flags)
1505 {
1506         struct vm_area_struct *vma;
1507         struct mm_struct *mm = current->mm;
1508         unsigned long addr = addr0, start_addr;
1509
1510         /* requested length too big for entire address space */
1511         if (len > TASK_SIZE)
1512                 return -ENOMEM;
1513
1514         if (flags & MAP_FIXED)
1515                 return addr;
1516
1517         /* requesting a specific address */
1518         if (addr) {
1519                 addr = PAGE_ALIGN(addr);
1520                 vma = find_vma(mm, addr);
1521                 if (TASK_SIZE - len >= addr &&
1522                                 (!vma || addr + len <= vma->vm_start))
1523                         return addr;
1524         }
1525
1526         /* check if free_area_cache is useful for us */
1527         if (len <= mm->cached_hole_size) {
1528                 mm->cached_hole_size = 0;
1529                 mm->free_area_cache = mm->mmap_base;
1530         }
1531
1532 try_again:
1533         /* either no address requested or can't fit in requested address hole */
1534         start_addr = addr = mm->free_area_cache;
1535
1536         if (addr < len)
1537                 goto fail;
1538
1539         addr -= len;
1540         do {
1541                 /*
1542                  * Lookup failure means no vma is above this address,
1543                  * else if new region fits below vma->vm_start,
1544                  * return with success:
1545                  */
1546                 vma = find_vma(mm, addr);
1547                 if (!vma || addr+len <= vma->vm_start)
1548                         /* remember the address as a hint for next time */
1549                         return (mm->free_area_cache = addr);
1550
1551                 /* remember the largest hole we saw so far */
1552                 if (addr + mm->cached_hole_size < vma->vm_start)
1553                         mm->cached_hole_size = vma->vm_start - addr;
1554
1555                 /* try just below the current vma->vm_start */
1556                 addr = vma->vm_start-len;
1557         } while (len < vma->vm_start);
1558
1559 fail:
1560         /*
1561          * if hint left us with no space for the requested
1562          * mapping then try again:
1563          *
1564          * Note: this is different with the case of bottomup
1565          * which does the fully line-search, but we use find_vma
1566          * here that causes some holes skipped.
1567          */
1568         if (start_addr != mm->mmap_base) {
1569                 mm->free_area_cache = mm->mmap_base;
1570                 mm->cached_hole_size = 0;
1571                 goto try_again;
1572         }
1573
1574         /*
1575          * A failed mmap() very likely causes application failure,
1576          * so fall back to the bottom-up function here. This scenario
1577          * can happen with large stack limits and large mmap()
1578          * allocations.
1579          */
1580         mm->cached_hole_size = ~0UL;
1581         mm->free_area_cache = TASK_UNMAPPED_BASE;
1582         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1583         /*
1584          * Restore the topdown base:
1585          */
1586         mm->free_area_cache = mm->mmap_base;
1587         mm->cached_hole_size = ~0UL;
1588
1589         return addr;
1590 }
1591 #endif
1592
1593 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1594 {
1595         /*
1596          * Is this a new hole at the highest possible address?
1597          */
1598         if (addr > mm->free_area_cache)
1599                 mm->free_area_cache = addr;
1600
1601         /* dont allow allocations above current base */
1602         if (mm->free_area_cache > mm->mmap_base)
1603                 mm->free_area_cache = mm->mmap_base;
1604 }
1605
1606 unsigned long
1607 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1608                 unsigned long pgoff, unsigned long flags)
1609 {
1610         unsigned long (*get_area)(struct file *, unsigned long,
1611                                   unsigned long, unsigned long, unsigned long);
1612
1613         unsigned long error = arch_mmap_check(addr, len, flags);
1614         if (error)
1615                 return error;
1616
1617         /* Careful about overflows.. */
1618         if (len > TASK_SIZE)
1619                 return -ENOMEM;
1620
1621         get_area = current->mm->get_unmapped_area;
1622         if (file && file->f_op && file->f_op->get_unmapped_area)
1623                 get_area = file->f_op->get_unmapped_area;
1624         addr = get_area(file, addr, len, pgoff, flags);
1625         if (IS_ERR_VALUE(addr))
1626                 return addr;
1627
1628         if (addr > TASK_SIZE - len)
1629                 return -ENOMEM;
1630         if (addr & ~PAGE_MASK)
1631                 return -EINVAL;
1632
1633         addr = arch_rebalance_pgtables(addr, len);
1634         error = security_mmap_addr(addr);
1635         return error ? error : addr;
1636 }
1637
1638 EXPORT_SYMBOL(get_unmapped_area);
1639
1640 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1641 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1642 {
1643         struct vm_area_struct *vma = NULL;
1644
1645         if (WARN_ON_ONCE(!mm))          /* Remove this in linux-3.6 */
1646                 return NULL;
1647
1648         /* Check the cache first. */
1649         /* (Cache hit rate is typically around 35%.) */
1650         vma = mm->mmap_cache;
1651         if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1652                 struct rb_node *rb_node;
1653
1654                 rb_node = mm->mm_rb.rb_node;
1655                 vma = NULL;
1656
1657                 while (rb_node) {
1658                         struct vm_area_struct *vma_tmp;
1659
1660                         vma_tmp = rb_entry(rb_node,
1661                                            struct vm_area_struct, vm_rb);
1662
1663                         if (vma_tmp->vm_end > addr) {
1664                                 vma = vma_tmp;
1665                                 if (vma_tmp->vm_start <= addr)
1666                                         break;
1667                                 rb_node = rb_node->rb_left;
1668                         } else
1669                                 rb_node = rb_node->rb_right;
1670                 }
1671                 if (vma)
1672                         mm->mmap_cache = vma;
1673         }
1674         return vma;
1675 }
1676
1677 EXPORT_SYMBOL(find_vma);
1678
1679 /*
1680  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1681  */
1682 struct vm_area_struct *
1683 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1684                         struct vm_area_struct **pprev)
1685 {
1686         struct vm_area_struct *vma;
1687
1688         vma = find_vma(mm, addr);
1689         if (vma) {
1690                 *pprev = vma->vm_prev;
1691         } else {
1692                 struct rb_node *rb_node = mm->mm_rb.rb_node;
1693                 *pprev = NULL;
1694                 while (rb_node) {
1695                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1696                         rb_node = rb_node->rb_right;
1697                 }
1698         }
1699         return vma;
1700 }
1701
1702 /*
1703  * Verify that the stack growth is acceptable and
1704  * update accounting. This is shared with both the
1705  * grow-up and grow-down cases.
1706  */
1707 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1708 {
1709         struct mm_struct *mm = vma->vm_mm;
1710         struct rlimit *rlim = current->signal->rlim;
1711         unsigned long new_start;
1712
1713         /* address space limit tests */
1714         if (!may_expand_vm(mm, grow))
1715                 return -ENOMEM;
1716
1717         /* Stack limit test */
1718         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1719                 return -ENOMEM;
1720
1721         /* mlock limit tests */
1722         if (vma->vm_flags & VM_LOCKED) {
1723                 unsigned long locked;
1724                 unsigned long limit;
1725                 locked = mm->locked_vm + grow;
1726                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1727                 limit >>= PAGE_SHIFT;
1728                 if (locked > limit && !capable(CAP_IPC_LOCK))
1729                         return -ENOMEM;
1730         }
1731
1732         /* Check to ensure the stack will not grow into a hugetlb-only region */
1733         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1734                         vma->vm_end - size;
1735         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1736                 return -EFAULT;
1737
1738         /*
1739          * Overcommit..  This must be the final test, as it will
1740          * update security statistics.
1741          */
1742         if (security_vm_enough_memory_mm(mm, grow))
1743                 return -ENOMEM;
1744
1745         /* Ok, everything looks good - let it rip */
1746         if (vma->vm_flags & VM_LOCKED)
1747                 mm->locked_vm += grow;
1748         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1749         return 0;
1750 }
1751
1752 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1753 /*
1754  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1755  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1756  */
1757 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1758 {
1759         int error;
1760
1761         if (!(vma->vm_flags & VM_GROWSUP))
1762                 return -EFAULT;
1763
1764         /*
1765          * We must make sure the anon_vma is allocated
1766          * so that the anon_vma locking is not a noop.
1767          */
1768         if (unlikely(anon_vma_prepare(vma)))
1769                 return -ENOMEM;
1770         vma_lock_anon_vma(vma);
1771
1772         /*
1773          * vma->vm_start/vm_end cannot change under us because the caller
1774          * is required to hold the mmap_sem in read mode.  We need the
1775          * anon_vma lock to serialize against concurrent expand_stacks.
1776          * Also guard against wrapping around to address 0.
1777          */
1778         if (address < PAGE_ALIGN(address+4))
1779                 address = PAGE_ALIGN(address+4);
1780         else {
1781                 vma_unlock_anon_vma(vma);
1782                 return -ENOMEM;
1783         }
1784         error = 0;
1785
1786         /* Somebody else might have raced and expanded it already */
1787         if (address > vma->vm_end) {
1788                 unsigned long size, grow;
1789
1790                 size = address - vma->vm_start;
1791                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1792
1793                 error = -ENOMEM;
1794                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1795                         error = acct_stack_growth(vma, size, grow);
1796                         if (!error) {
1797                                 anon_vma_interval_tree_pre_update_vma(vma);
1798                                 vma->vm_end = address;
1799                                 anon_vma_interval_tree_post_update_vma(vma);
1800                                 perf_event_mmap(vma);
1801                         }
1802                 }
1803         }
1804         vma_unlock_anon_vma(vma);
1805         khugepaged_enter_vma_merge(vma);
1806         validate_mm(vma->vm_mm);
1807         return error;
1808 }
1809 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1810
1811 /*
1812  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1813  */
1814 int expand_downwards(struct vm_area_struct *vma,
1815                                    unsigned long address)
1816 {
1817         int error;
1818
1819         /*
1820          * We must make sure the anon_vma is allocated
1821          * so that the anon_vma locking is not a noop.
1822          */
1823         if (unlikely(anon_vma_prepare(vma)))
1824                 return -ENOMEM;
1825
1826         address &= PAGE_MASK;
1827         error = security_mmap_addr(address);
1828         if (error)
1829                 return error;
1830
1831         vma_lock_anon_vma(vma);
1832
1833         /*
1834          * vma->vm_start/vm_end cannot change under us because the caller
1835          * is required to hold the mmap_sem in read mode.  We need the
1836          * anon_vma lock to serialize against concurrent expand_stacks.
1837          */
1838
1839         /* Somebody else might have raced and expanded it already */
1840         if (address < vma->vm_start) {
1841                 unsigned long size, grow;
1842
1843                 size = vma->vm_end - address;
1844                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1845
1846                 error = -ENOMEM;
1847                 if (grow <= vma->vm_pgoff) {
1848                         error = acct_stack_growth(vma, size, grow);
1849                         if (!error) {
1850                                 anon_vma_interval_tree_pre_update_vma(vma);
1851                                 vma->vm_start = address;
1852                                 vma->vm_pgoff -= grow;
1853                                 anon_vma_interval_tree_post_update_vma(vma);
1854                                 perf_event_mmap(vma);
1855                         }
1856                 }
1857         }
1858         vma_unlock_anon_vma(vma);
1859         khugepaged_enter_vma_merge(vma);
1860         validate_mm(vma->vm_mm);
1861         return error;
1862 }
1863
1864 #ifdef CONFIG_STACK_GROWSUP
1865 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1866 {
1867         return expand_upwards(vma, address);
1868 }
1869
1870 struct vm_area_struct *
1871 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1872 {
1873         struct vm_area_struct *vma, *prev;
1874
1875         addr &= PAGE_MASK;
1876         vma = find_vma_prev(mm, addr, &prev);
1877         if (vma && (vma->vm_start <= addr))
1878                 return vma;
1879         if (!prev || expand_stack(prev, addr))
1880                 return NULL;
1881         if (prev->vm_flags & VM_LOCKED) {
1882                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1883         }
1884         return prev;
1885 }
1886 #else
1887 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1888 {
1889         return expand_downwards(vma, address);
1890 }
1891
1892 struct vm_area_struct *
1893 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1894 {
1895         struct vm_area_struct * vma;
1896         unsigned long start;
1897
1898         addr &= PAGE_MASK;
1899         vma = find_vma(mm,addr);
1900         if (!vma)
1901                 return NULL;
1902         if (vma->vm_start <= addr)
1903                 return vma;
1904         if (!(vma->vm_flags & VM_GROWSDOWN))
1905                 return NULL;
1906         start = vma->vm_start;
1907         if (expand_stack(vma, addr))
1908                 return NULL;
1909         if (vma->vm_flags & VM_LOCKED) {
1910                 mlock_vma_pages_range(vma, addr, start);
1911         }
1912         return vma;
1913 }
1914 #endif
1915
1916 /*
1917  * Ok - we have the memory areas we should free on the vma list,
1918  * so release them, and do the vma updates.
1919  *
1920  * Called with the mm semaphore held.
1921  */
1922 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1923 {
1924         unsigned long nr_accounted = 0;
1925
1926         /* Update high watermark before we lower total_vm */
1927         update_hiwater_vm(mm);
1928         do {
1929                 long nrpages = vma_pages(vma);
1930
1931                 if (vma->vm_flags & VM_ACCOUNT)
1932                         nr_accounted += nrpages;
1933                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1934                 vma = remove_vma(vma);
1935         } while (vma);
1936         vm_unacct_memory(nr_accounted);
1937         validate_mm(mm);
1938 }
1939
1940 /*
1941  * Get rid of page table information in the indicated region.
1942  *
1943  * Called with the mm semaphore held.
1944  */
1945 static void unmap_region(struct mm_struct *mm,
1946                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1947                 unsigned long start, unsigned long end)
1948 {
1949         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1950         struct mmu_gather tlb;
1951
1952         lru_add_drain();
1953         tlb_gather_mmu(&tlb, mm, 0);
1954         update_hiwater_rss(mm);
1955         unmap_vmas(&tlb, vma, start, end);
1956         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1957                                  next ? next->vm_start : 0);
1958         tlb_finish_mmu(&tlb, start, end);
1959 }
1960
1961 /*
1962  * Create a list of vma's touched by the unmap, removing them from the mm's
1963  * vma list as we go..
1964  */
1965 static void
1966 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1967         struct vm_area_struct *prev, unsigned long end)
1968 {
1969         struct vm_area_struct **insertion_point;
1970         struct vm_area_struct *tail_vma = NULL;
1971         unsigned long addr;
1972
1973         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1974         vma->vm_prev = NULL;
1975         do {
1976                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1977                 mm->map_count--;
1978                 tail_vma = vma;
1979                 vma = vma->vm_next;
1980         } while (vma && vma->vm_start < end);
1981         *insertion_point = vma;
1982         if (vma)
1983                 vma->vm_prev = prev;
1984         tail_vma->vm_next = NULL;
1985         if (mm->unmap_area == arch_unmap_area)
1986                 addr = prev ? prev->vm_end : mm->mmap_base;
1987         else
1988                 addr = vma ?  vma->vm_start : mm->mmap_base;
1989         mm->unmap_area(mm, addr);
1990         mm->mmap_cache = NULL;          /* Kill the cache. */
1991 }
1992
1993 /*
1994  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1995  * munmap path where it doesn't make sense to fail.
1996  */
1997 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1998               unsigned long addr, int new_below)
1999 {
2000         struct mempolicy *pol;
2001         struct vm_area_struct *new;
2002         int err = -ENOMEM;
2003
2004         if (is_vm_hugetlb_page(vma) && (addr &
2005                                         ~(huge_page_mask(hstate_vma(vma)))))
2006                 return -EINVAL;
2007
2008         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2009         if (!new)
2010                 goto out_err;
2011
2012         /* most fields are the same, copy all, and then fixup */
2013         *new = *vma;
2014
2015         INIT_LIST_HEAD(&new->anon_vma_chain);
2016
2017         if (new_below)
2018                 new->vm_end = addr;
2019         else {
2020                 new->vm_start = addr;
2021                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2022         }
2023
2024         pol = mpol_dup(vma_policy(vma));
2025         if (IS_ERR(pol)) {
2026                 err = PTR_ERR(pol);
2027                 goto out_free_vma;
2028         }
2029         vma_set_policy(new, pol);
2030
2031         if (anon_vma_clone(new, vma))
2032                 goto out_free_mpol;
2033
2034         if (new->vm_file)
2035                 get_file(new->vm_file);
2036
2037         if (new->vm_ops && new->vm_ops->open)
2038                 new->vm_ops->open(new);
2039
2040         if (new_below)
2041                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2042                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2043         else
2044                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2045
2046         /* Success. */
2047         if (!err)
2048                 return 0;
2049
2050         /* Clean everything up if vma_adjust failed. */
2051         if (new->vm_ops && new->vm_ops->close)
2052                 new->vm_ops->close(new);
2053         if (new->vm_file)
2054                 fput(new->vm_file);
2055         unlink_anon_vmas(new);
2056  out_free_mpol:
2057         mpol_put(pol);
2058  out_free_vma:
2059         kmem_cache_free(vm_area_cachep, new);
2060  out_err:
2061         return err;
2062 }
2063
2064 /*
2065  * Split a vma into two pieces at address 'addr', a new vma is allocated
2066  * either for the first part or the tail.
2067  */
2068 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2069               unsigned long addr, int new_below)
2070 {
2071         if (mm->map_count >= sysctl_max_map_count)
2072                 return -ENOMEM;
2073
2074         return __split_vma(mm, vma, addr, new_below);
2075 }
2076
2077 /* Munmap is split into 2 main parts -- this part which finds
2078  * what needs doing, and the areas themselves, which do the
2079  * work.  This now handles partial unmappings.
2080  * Jeremy Fitzhardinge <jeremy@goop.org>
2081  */
2082 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2083 {
2084         unsigned long end;
2085         struct vm_area_struct *vma, *prev, *last;
2086
2087         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2088                 return -EINVAL;
2089
2090         if ((len = PAGE_ALIGN(len)) == 0)
2091                 return -EINVAL;
2092
2093         /* Find the first overlapping VMA */
2094         vma = find_vma(mm, start);
2095         if (!vma)
2096                 return 0;
2097         prev = vma->vm_prev;
2098         /* we have  start < vma->vm_end  */
2099
2100         /* if it doesn't overlap, we have nothing.. */
2101         end = start + len;
2102         if (vma->vm_start >= end)
2103                 return 0;
2104
2105         /*
2106          * If we need to split any vma, do it now to save pain later.
2107          *
2108          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2109          * unmapped vm_area_struct will remain in use: so lower split_vma
2110          * places tmp vma above, and higher split_vma places tmp vma below.
2111          */
2112         if (start > vma->vm_start) {
2113                 int error;
2114
2115                 /*
2116                  * Make sure that map_count on return from munmap() will
2117                  * not exceed its limit; but let map_count go just above
2118                  * its limit temporarily, to help free resources as expected.
2119                  */
2120                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2121                         return -ENOMEM;
2122
2123                 error = __split_vma(mm, vma, start, 0);
2124                 if (error)
2125                         return error;
2126                 prev = vma;
2127         }
2128
2129         /* Does it split the last one? */
2130         last = find_vma(mm, end);
2131         if (last && end > last->vm_start) {
2132                 int error = __split_vma(mm, last, end, 1);
2133                 if (error)
2134                         return error;
2135         }
2136         vma = prev? prev->vm_next: mm->mmap;
2137
2138         /*
2139          * unlock any mlock()ed ranges before detaching vmas
2140          */
2141         if (mm->locked_vm) {
2142                 struct vm_area_struct *tmp = vma;
2143                 while (tmp && tmp->vm_start < end) {
2144                         if (tmp->vm_flags & VM_LOCKED) {
2145                                 mm->locked_vm -= vma_pages(tmp);
2146                                 munlock_vma_pages_all(tmp);
2147                         }
2148                         tmp = tmp->vm_next;
2149                 }
2150         }
2151
2152         /*
2153          * Remove the vma's, and unmap the actual pages
2154          */
2155         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2156         unmap_region(mm, vma, prev, start, end);
2157
2158         /* Fix up all other VM information */
2159         remove_vma_list(mm, vma);
2160
2161         return 0;
2162 }
2163
2164 int vm_munmap(unsigned long start, size_t len)
2165 {
2166         int ret;
2167         struct mm_struct *mm = current->mm;
2168
2169         down_write(&mm->mmap_sem);
2170         ret = do_munmap(mm, start, len);
2171         up_write(&mm->mmap_sem);
2172         return ret;
2173 }
2174 EXPORT_SYMBOL(vm_munmap);
2175
2176 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2177 {
2178         profile_munmap(addr);
2179         return vm_munmap(addr, len);
2180 }
2181
2182 static inline void verify_mm_writelocked(struct mm_struct *mm)
2183 {
2184 #ifdef CONFIG_DEBUG_VM
2185         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2186                 WARN_ON(1);
2187                 up_read(&mm->mmap_sem);
2188         }
2189 #endif
2190 }
2191
2192 /*
2193  *  this is really a simplified "do_mmap".  it only handles
2194  *  anonymous maps.  eventually we may be able to do some
2195  *  brk-specific accounting here.
2196  */
2197 static unsigned long do_brk(unsigned long addr, unsigned long len)
2198 {
2199         struct mm_struct * mm = current->mm;
2200         struct vm_area_struct * vma, * prev;
2201         unsigned long flags;
2202         struct rb_node ** rb_link, * rb_parent;
2203         pgoff_t pgoff = addr >> PAGE_SHIFT;
2204         int error;
2205
2206         len = PAGE_ALIGN(len);
2207         if (!len)
2208                 return addr;
2209
2210         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2211
2212         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2213         if (error & ~PAGE_MASK)
2214                 return error;
2215
2216         /*
2217          * mlock MCL_FUTURE?
2218          */
2219         if (mm->def_flags & VM_LOCKED) {
2220                 unsigned long locked, lock_limit;
2221                 locked = len >> PAGE_SHIFT;
2222                 locked += mm->locked_vm;
2223                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2224                 lock_limit >>= PAGE_SHIFT;
2225                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2226                         return -EAGAIN;
2227         }
2228
2229         /*
2230          * mm->mmap_sem is required to protect against another thread
2231          * changing the mappings in case we sleep.
2232          */
2233         verify_mm_writelocked(mm);
2234
2235         /*
2236          * Clear old maps.  this also does some error checking for us
2237          */
2238  munmap_back:
2239         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
2240                 if (do_munmap(mm, addr, len))
2241                         return -ENOMEM;
2242                 goto munmap_back;
2243         }
2244
2245         /* Check against address space limits *after* clearing old maps... */
2246         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2247                 return -ENOMEM;
2248
2249         if (mm->map_count > sysctl_max_map_count)
2250                 return -ENOMEM;
2251
2252         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2253                 return -ENOMEM;
2254
2255         /* Can we just expand an old private anonymous mapping? */
2256         vma = vma_merge(mm, prev, addr, addr + len, flags,
2257                                         NULL, NULL, pgoff, NULL);
2258         if (vma)
2259                 goto out;
2260
2261         /*
2262          * create a vma struct for an anonymous mapping
2263          */
2264         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2265         if (!vma) {
2266                 vm_unacct_memory(len >> PAGE_SHIFT);
2267                 return -ENOMEM;
2268         }
2269
2270         INIT_LIST_HEAD(&vma->anon_vma_chain);
2271         vma->vm_mm = mm;
2272         vma->vm_start = addr;
2273         vma->vm_end = addr + len;
2274         vma->vm_pgoff = pgoff;
2275         vma->vm_flags = flags;
2276         vma->vm_page_prot = vm_get_page_prot(flags);
2277         vma_link(mm, vma, prev, rb_link, rb_parent);
2278 out:
2279         perf_event_mmap(vma);
2280         mm->total_vm += len >> PAGE_SHIFT;
2281         if (flags & VM_LOCKED) {
2282                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2283                         mm->locked_vm += (len >> PAGE_SHIFT);
2284         }
2285         return addr;
2286 }
2287
2288 unsigned long vm_brk(unsigned long addr, unsigned long len)
2289 {
2290         struct mm_struct *mm = current->mm;
2291         unsigned long ret;
2292
2293         down_write(&mm->mmap_sem);
2294         ret = do_brk(addr, len);
2295         up_write(&mm->mmap_sem);
2296         return ret;
2297 }
2298 EXPORT_SYMBOL(vm_brk);
2299
2300 /* Release all mmaps. */
2301 void exit_mmap(struct mm_struct *mm)
2302 {
2303         struct mmu_gather tlb;
2304         struct vm_area_struct *vma;
2305         unsigned long nr_accounted = 0;
2306
2307         /* mm's last user has gone, and its about to be pulled down */
2308         mmu_notifier_release(mm);
2309
2310         if (mm->locked_vm) {
2311                 vma = mm->mmap;
2312                 while (vma) {
2313                         if (vma->vm_flags & VM_LOCKED)
2314                                 munlock_vma_pages_all(vma);
2315                         vma = vma->vm_next;
2316                 }
2317         }
2318
2319         arch_exit_mmap(mm);
2320
2321         vma = mm->mmap;
2322         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2323                 return;
2324
2325         lru_add_drain();
2326         flush_cache_mm(mm);
2327         tlb_gather_mmu(&tlb, mm, 1);
2328         /* update_hiwater_rss(mm) here? but nobody should be looking */
2329         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2330         unmap_vmas(&tlb, vma, 0, -1);
2331
2332         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2333         tlb_finish_mmu(&tlb, 0, -1);
2334
2335         /*
2336          * Walk the list again, actually closing and freeing it,
2337          * with preemption enabled, without holding any MM locks.
2338          */
2339         while (vma) {
2340                 if (vma->vm_flags & VM_ACCOUNT)
2341                         nr_accounted += vma_pages(vma);
2342                 vma = remove_vma(vma);
2343         }
2344         vm_unacct_memory(nr_accounted);
2345
2346         WARN_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2347 }
2348
2349 /* Insert vm structure into process list sorted by address
2350  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2351  * then i_mmap_mutex is taken here.
2352  */
2353 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2354 {
2355         struct vm_area_struct *prev;
2356         struct rb_node **rb_link, *rb_parent;
2357
2358         /*
2359          * The vm_pgoff of a purely anonymous vma should be irrelevant
2360          * until its first write fault, when page's anon_vma and index
2361          * are set.  But now set the vm_pgoff it will almost certainly
2362          * end up with (unless mremap moves it elsewhere before that
2363          * first wfault), so /proc/pid/maps tells a consistent story.
2364          *
2365          * By setting it to reflect the virtual start address of the
2366          * vma, merges and splits can happen in a seamless way, just
2367          * using the existing file pgoff checks and manipulations.
2368          * Similarly in do_mmap_pgoff and in do_brk.
2369          */
2370         if (!vma->vm_file) {
2371                 BUG_ON(vma->anon_vma);
2372                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2373         }
2374         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2375                            &prev, &rb_link, &rb_parent))
2376                 return -ENOMEM;
2377         if ((vma->vm_flags & VM_ACCOUNT) &&
2378              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2379                 return -ENOMEM;
2380
2381         vma_link(mm, vma, prev, rb_link, rb_parent);
2382         return 0;
2383 }
2384
2385 /*
2386  * Copy the vma structure to a new location in the same mm,
2387  * prior to moving page table entries, to effect an mremap move.
2388  */
2389 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2390         unsigned long addr, unsigned long len, pgoff_t pgoff,
2391         bool *need_rmap_locks)
2392 {
2393         struct vm_area_struct *vma = *vmap;
2394         unsigned long vma_start = vma->vm_start;
2395         struct mm_struct *mm = vma->vm_mm;
2396         struct vm_area_struct *new_vma, *prev;
2397         struct rb_node **rb_link, *rb_parent;
2398         struct mempolicy *pol;
2399         bool faulted_in_anon_vma = true;
2400
2401         /*
2402          * If anonymous vma has not yet been faulted, update new pgoff
2403          * to match new location, to increase its chance of merging.
2404          */
2405         if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2406                 pgoff = addr >> PAGE_SHIFT;
2407                 faulted_in_anon_vma = false;
2408         }
2409
2410         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2411                 return NULL;    /* should never get here */
2412         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2413                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2414         if (new_vma) {
2415                 /*
2416                  * Source vma may have been merged into new_vma
2417                  */
2418                 if (unlikely(vma_start >= new_vma->vm_start &&
2419                              vma_start < new_vma->vm_end)) {
2420                         /*
2421                          * The only way we can get a vma_merge with
2422                          * self during an mremap is if the vma hasn't
2423                          * been faulted in yet and we were allowed to
2424                          * reset the dst vma->vm_pgoff to the
2425                          * destination address of the mremap to allow
2426                          * the merge to happen. mremap must change the
2427                          * vm_pgoff linearity between src and dst vmas
2428                          * (in turn preventing a vma_merge) to be
2429                          * safe. It is only safe to keep the vm_pgoff
2430                          * linear if there are no pages mapped yet.
2431                          */
2432                         VM_BUG_ON(faulted_in_anon_vma);
2433                         *vmap = vma = new_vma;
2434                 }
2435                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2436         } else {
2437                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2438                 if (new_vma) {
2439                         *new_vma = *vma;
2440                         new_vma->vm_start = addr;
2441                         new_vma->vm_end = addr + len;
2442                         new_vma->vm_pgoff = pgoff;
2443                         pol = mpol_dup(vma_policy(vma));
2444                         if (IS_ERR(pol))
2445                                 goto out_free_vma;
2446                         vma_set_policy(new_vma, pol);
2447                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2448                         if (anon_vma_clone(new_vma, vma))
2449                                 goto out_free_mempol;
2450                         if (new_vma->vm_file)
2451                                 get_file(new_vma->vm_file);
2452                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2453                                 new_vma->vm_ops->open(new_vma);
2454                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2455                         *need_rmap_locks = false;
2456                 }
2457         }
2458         return new_vma;
2459
2460  out_free_mempol:
2461         mpol_put(pol);
2462  out_free_vma:
2463         kmem_cache_free(vm_area_cachep, new_vma);
2464         return NULL;
2465 }
2466
2467 /*
2468  * Return true if the calling process may expand its vm space by the passed
2469  * number of pages
2470  */
2471 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2472 {
2473         unsigned long cur = mm->total_vm;       /* pages */
2474         unsigned long lim;
2475
2476         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2477
2478         if (cur + npages > lim)
2479                 return 0;
2480         return 1;
2481 }
2482
2483
2484 static int special_mapping_fault(struct vm_area_struct *vma,
2485                                 struct vm_fault *vmf)
2486 {
2487         pgoff_t pgoff;
2488         struct page **pages;
2489
2490         /*
2491          * special mappings have no vm_file, and in that case, the mm
2492          * uses vm_pgoff internally. So we have to subtract it from here.
2493          * We are allowed to do this because we are the mm; do not copy
2494          * this code into drivers!
2495          */
2496         pgoff = vmf->pgoff - vma->vm_pgoff;
2497
2498         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2499                 pgoff--;
2500
2501         if (*pages) {
2502                 struct page *page = *pages;
2503                 get_page(page);
2504                 vmf->page = page;
2505                 return 0;
2506         }
2507
2508         return VM_FAULT_SIGBUS;
2509 }
2510
2511 /*
2512  * Having a close hook prevents vma merging regardless of flags.
2513  */
2514 static void special_mapping_close(struct vm_area_struct *vma)
2515 {
2516 }
2517
2518 static const struct vm_operations_struct special_mapping_vmops = {
2519         .close = special_mapping_close,
2520         .fault = special_mapping_fault,
2521 };
2522
2523 /*
2524  * Called with mm->mmap_sem held for writing.
2525  * Insert a new vma covering the given region, with the given flags.
2526  * Its pages are supplied by the given array of struct page *.
2527  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2528  * The region past the last page supplied will always produce SIGBUS.
2529  * The array pointer and the pages it points to are assumed to stay alive
2530  * for as long as this mapping might exist.
2531  */
2532 int install_special_mapping(struct mm_struct *mm,
2533                             unsigned long addr, unsigned long len,
2534                             unsigned long vm_flags, struct page **pages)
2535 {
2536         int ret;
2537         struct vm_area_struct *vma;
2538
2539         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2540         if (unlikely(vma == NULL))
2541                 return -ENOMEM;
2542
2543         INIT_LIST_HEAD(&vma->anon_vma_chain);
2544         vma->vm_mm = mm;
2545         vma->vm_start = addr;
2546         vma->vm_end = addr + len;
2547
2548         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2549         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2550
2551         vma->vm_ops = &special_mapping_vmops;
2552         vma->vm_private_data = pages;
2553
2554         ret = insert_vm_struct(mm, vma);
2555         if (ret)
2556                 goto out;
2557
2558         mm->total_vm += len >> PAGE_SHIFT;
2559
2560         perf_event_mmap(vma);
2561
2562         return 0;
2563
2564 out:
2565         kmem_cache_free(vm_area_cachep, vma);
2566         return ret;
2567 }
2568
2569 static DEFINE_MUTEX(mm_all_locks_mutex);
2570
2571 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2572 {
2573         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2574                 /*
2575                  * The LSB of head.next can't change from under us
2576                  * because we hold the mm_all_locks_mutex.
2577                  */
2578                 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2579                 /*
2580                  * We can safely modify head.next after taking the
2581                  * anon_vma->root->mutex. If some other vma in this mm shares
2582                  * the same anon_vma we won't take it again.
2583                  *
2584                  * No need of atomic instructions here, head.next
2585                  * can't change from under us thanks to the
2586                  * anon_vma->root->mutex.
2587                  */
2588                 if (__test_and_set_bit(0, (unsigned long *)
2589                                        &anon_vma->root->rb_root.rb_node))
2590                         BUG();
2591         }
2592 }
2593
2594 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2595 {
2596         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2597                 /*
2598                  * AS_MM_ALL_LOCKS can't change from under us because
2599                  * we hold the mm_all_locks_mutex.
2600                  *
2601                  * Operations on ->flags have to be atomic because
2602                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2603                  * mm_all_locks_mutex, there may be other cpus
2604                  * changing other bitflags in parallel to us.
2605                  */
2606                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2607                         BUG();
2608                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2609         }
2610 }
2611
2612 /*
2613  * This operation locks against the VM for all pte/vma/mm related
2614  * operations that could ever happen on a certain mm. This includes
2615  * vmtruncate, try_to_unmap, and all page faults.
2616  *
2617  * The caller must take the mmap_sem in write mode before calling
2618  * mm_take_all_locks(). The caller isn't allowed to release the
2619  * mmap_sem until mm_drop_all_locks() returns.
2620  *
2621  * mmap_sem in write mode is required in order to block all operations
2622  * that could modify pagetables and free pages without need of
2623  * altering the vma layout (for example populate_range() with
2624  * nonlinear vmas). It's also needed in write mode to avoid new
2625  * anon_vmas to be associated with existing vmas.
2626  *
2627  * A single task can't take more than one mm_take_all_locks() in a row
2628  * or it would deadlock.
2629  *
2630  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
2631  * mapping->flags avoid to take the same lock twice, if more than one
2632  * vma in this mm is backed by the same anon_vma or address_space.
2633  *
2634  * We can take all the locks in random order because the VM code
2635  * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2636  * takes more than one of them in a row. Secondly we're protected
2637  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2638  *
2639  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2640  * that may have to take thousand of locks.
2641  *
2642  * mm_take_all_locks() can fail if it's interrupted by signals.
2643  */
2644 int mm_take_all_locks(struct mm_struct *mm)
2645 {
2646         struct vm_area_struct *vma;
2647         struct anon_vma_chain *avc;
2648
2649         BUG_ON(down_read_trylock(&mm->mmap_sem));
2650
2651         mutex_lock(&mm_all_locks_mutex);
2652
2653         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2654                 if (signal_pending(current))
2655                         goto out_unlock;
2656                 if (vma->vm_file && vma->vm_file->f_mapping)
2657                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2658         }
2659
2660         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2661                 if (signal_pending(current))
2662                         goto out_unlock;
2663                 if (vma->anon_vma)
2664                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2665                                 vm_lock_anon_vma(mm, avc->anon_vma);
2666         }
2667
2668         return 0;
2669
2670 out_unlock:
2671         mm_drop_all_locks(mm);
2672         return -EINTR;
2673 }
2674
2675 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2676 {
2677         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2678                 /*
2679                  * The LSB of head.next can't change to 0 from under
2680                  * us because we hold the mm_all_locks_mutex.
2681                  *
2682                  * We must however clear the bitflag before unlocking
2683                  * the vma so the users using the anon_vma->rb_root will
2684                  * never see our bitflag.
2685                  *
2686                  * No need of atomic instructions here, head.next
2687                  * can't change from under us until we release the
2688                  * anon_vma->root->mutex.
2689                  */
2690                 if (!__test_and_clear_bit(0, (unsigned long *)
2691                                           &anon_vma->root->rb_root.rb_node))
2692                         BUG();
2693                 anon_vma_unlock(anon_vma);
2694         }
2695 }
2696
2697 static void vm_unlock_mapping(struct address_space *mapping)
2698 {
2699         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2700                 /*
2701                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2702                  * because we hold the mm_all_locks_mutex.
2703                  */
2704                 mutex_unlock(&mapping->i_mmap_mutex);
2705                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2706                                         &mapping->flags))
2707                         BUG();
2708         }
2709 }
2710
2711 /*
2712  * The mmap_sem cannot be released by the caller until
2713  * mm_drop_all_locks() returns.
2714  */
2715 void mm_drop_all_locks(struct mm_struct *mm)
2716 {
2717         struct vm_area_struct *vma;
2718         struct anon_vma_chain *avc;
2719
2720         BUG_ON(down_read_trylock(&mm->mmap_sem));
2721         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2722
2723         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2724                 if (vma->anon_vma)
2725                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2726                                 vm_unlock_anon_vma(avc->anon_vma);
2727                 if (vma->vm_file && vma->vm_file->f_mapping)
2728                         vm_unlock_mapping(vma->vm_file->f_mapping);
2729         }
2730
2731         mutex_unlock(&mm_all_locks_mutex);
2732 }
2733
2734 /*
2735  * initialise the VMA slab
2736  */
2737 void __init mmap_init(void)
2738 {
2739         int ret;
2740
2741         ret = percpu_counter_init(&vm_committed_as, 0);
2742         VM_BUG_ON(ret);
2743 }