]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - mm/mmap.c
Merge tag 'clk-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[karo-tx-linux.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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
14 #include <linux/mm.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
24 #include <linux/fs.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/profile.h>
29 #include <linux/export.h>
30 #include <linux/mount.h>
31 #include <linux/mempolicy.h>
32 #include <linux/rmap.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/mmdebug.h>
35 #include <linux/perf_event.h>
36 #include <linux/audit.h>
37 #include <linux/khugepaged.h>
38 #include <linux/uprobes.h>
39 #include <linux/rbtree_augmented.h>
40 #include <linux/sched/sysctl.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 #include <linux/userfaultfd_k.h>
45
46 #include <asm/uaccess.h>
47 #include <asm/cacheflush.h>
48 #include <asm/tlb.h>
49 #include <asm/mmu_context.h>
50
51 #include "internal.h"
52
53 #ifndef arch_mmap_check
54 #define arch_mmap_check(addr, len, flags)       (0)
55 #endif
56
57 #ifndef arch_rebalance_pgtables
58 #define arch_rebalance_pgtables(addr, len)              (addr)
59 #endif
60
61 static void unmap_region(struct mm_struct *mm,
62                 struct vm_area_struct *vma, struct vm_area_struct *prev,
63                 unsigned long start, unsigned long end);
64
65 /* description of effects of mapping type and prot in current implementation.
66  * this is due to the limited x86 page protection hardware.  The expected
67  * behavior is in parens:
68  *
69  * map_type     prot
70  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
71  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
72  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
73  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
74  *
75  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
76  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
77  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
78  *
79  */
80 pgprot_t protection_map[16] = {
81         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
82         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
83 };
84
85 pgprot_t vm_get_page_prot(unsigned long vm_flags)
86 {
87         return __pgprot(pgprot_val(protection_map[vm_flags &
88                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
89                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
90 }
91 EXPORT_SYMBOL(vm_get_page_prot);
92
93 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
94 {
95         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
96 }
97
98 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
99 void vma_set_page_prot(struct vm_area_struct *vma)
100 {
101         unsigned long vm_flags = vma->vm_flags;
102
103         vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
104         if (vma_wants_writenotify(vma)) {
105                 vm_flags &= ~VM_SHARED;
106                 vma->vm_page_prot = vm_pgprot_modify(vma->vm_page_prot,
107                                                      vm_flags);
108         }
109 }
110
111
112 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
113 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
114 unsigned long sysctl_overcommit_kbytes __read_mostly;
115 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
116 unsigned long sysctl_user_reserve_kbytes __read_mostly = 1UL << 17; /* 128MB */
117 unsigned long sysctl_admin_reserve_kbytes __read_mostly = 1UL << 13; /* 8MB */
118 /*
119  * Make sure vm_committed_as in one cacheline and not cacheline shared with
120  * other variables. It can be updated by several CPUs frequently.
121  */
122 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
123
124 /*
125  * The global memory commitment made in the system can be a metric
126  * that can be used to drive ballooning decisions when Linux is hosted
127  * as a guest. On Hyper-V, the host implements a policy engine for dynamically
128  * balancing memory across competing virtual machines that are hosted.
129  * Several metrics drive this policy engine including the guest reported
130  * memory commitment.
131  */
132 unsigned long vm_memory_committed(void)
133 {
134         return percpu_counter_read_positive(&vm_committed_as);
135 }
136 EXPORT_SYMBOL_GPL(vm_memory_committed);
137
138 /*
139  * Check that a process has enough memory to allocate a new virtual
140  * mapping. 0 means there is enough memory for the allocation to
141  * succeed and -ENOMEM implies there is not.
142  *
143  * We currently support three overcommit policies, which are set via the
144  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
145  *
146  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
147  * Additional code 2002 Jul 20 by Robert Love.
148  *
149  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
150  *
151  * Note this is a helper function intended to be used by LSMs which
152  * wish to use this logic.
153  */
154 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
155 {
156         long free, allowed, reserve;
157
158         VM_WARN_ONCE(percpu_counter_read(&vm_committed_as) <
159                         -(s64)vm_committed_as_batch * num_online_cpus(),
160                         "memory commitment underflow");
161
162         vm_acct_memory(pages);
163
164         /*
165          * Sometimes we want to use more memory than we have
166          */
167         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
168                 return 0;
169
170         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
171                 free = global_page_state(NR_FREE_PAGES);
172                 free += global_page_state(NR_FILE_PAGES);
173
174                 /*
175                  * shmem pages shouldn't be counted as free in this
176                  * case, they can't be purged, only swapped out, and
177                  * that won't affect the overall amount of available
178                  * memory in the system.
179                  */
180                 free -= global_page_state(NR_SHMEM);
181
182                 free += get_nr_swap_pages();
183
184                 /*
185                  * Any slabs which are created with the
186                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
187                  * which are reclaimable, under pressure.  The dentry
188                  * cache and most inode caches should fall into this
189                  */
190                 free += global_page_state(NR_SLAB_RECLAIMABLE);
191
192                 /*
193                  * Leave reserved pages. The pages are not for anonymous pages.
194                  */
195                 if (free <= totalreserve_pages)
196                         goto error;
197                 else
198                         free -= totalreserve_pages;
199
200                 /*
201                  * Reserve some for root
202                  */
203                 if (!cap_sys_admin)
204                         free -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
205
206                 if (free > pages)
207                         return 0;
208
209                 goto error;
210         }
211
212         allowed = vm_commit_limit();
213         /*
214          * Reserve some for root
215          */
216         if (!cap_sys_admin)
217                 allowed -= sysctl_admin_reserve_kbytes >> (PAGE_SHIFT - 10);
218
219         /*
220          * Don't let a single process grow so big a user can't recover
221          */
222         if (mm) {
223                 reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
224                 allowed -= min_t(long, mm->total_vm / 32, reserve);
225         }
226
227         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
228                 return 0;
229 error:
230         vm_unacct_memory(pages);
231
232         return -ENOMEM;
233 }
234
235 /*
236  * Requires inode->i_mapping->i_mmap_rwsem
237  */
238 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
239                 struct file *file, struct address_space *mapping)
240 {
241         if (vma->vm_flags & VM_DENYWRITE)
242                 atomic_inc(&file_inode(file)->i_writecount);
243         if (vma->vm_flags & VM_SHARED)
244                 mapping_unmap_writable(mapping);
245
246         flush_dcache_mmap_lock(mapping);
247         vma_interval_tree_remove(vma, &mapping->i_mmap);
248         flush_dcache_mmap_unlock(mapping);
249 }
250
251 /*
252  * Unlink a file-based vm structure from its interval tree, to hide
253  * vma from rmap and vmtruncate before freeing its page tables.
254  */
255 void unlink_file_vma(struct vm_area_struct *vma)
256 {
257         struct file *file = vma->vm_file;
258
259         if (file) {
260                 struct address_space *mapping = file->f_mapping;
261                 i_mmap_lock_write(mapping);
262                 __remove_shared_vm_struct(vma, file, mapping);
263                 i_mmap_unlock_write(mapping);
264         }
265 }
266
267 /*
268  * Close a vm structure and free it, returning the next.
269  */
270 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
271 {
272         struct vm_area_struct *next = vma->vm_next;
273
274         might_sleep();
275         if (vma->vm_ops && vma->vm_ops->close)
276                 vma->vm_ops->close(vma);
277         if (vma->vm_file)
278                 fput(vma->vm_file);
279         mpol_put(vma_policy(vma));
280         kmem_cache_free(vm_area_cachep, vma);
281         return next;
282 }
283
284 static unsigned long do_brk(unsigned long addr, unsigned long len);
285
286 SYSCALL_DEFINE1(brk, unsigned long, brk)
287 {
288         unsigned long retval;
289         unsigned long newbrk, oldbrk;
290         struct mm_struct *mm = current->mm;
291         unsigned long min_brk;
292         bool populate;
293
294         down_write(&mm->mmap_sem);
295
296 #ifdef CONFIG_COMPAT_BRK
297         /*
298          * CONFIG_COMPAT_BRK can still be overridden by setting
299          * randomize_va_space to 2, which will still cause mm->start_brk
300          * to be arbitrarily shifted
301          */
302         if (current->brk_randomized)
303                 min_brk = mm->start_brk;
304         else
305                 min_brk = mm->end_data;
306 #else
307         min_brk = mm->start_brk;
308 #endif
309         if (brk < min_brk)
310                 goto out;
311
312         /*
313          * Check against rlimit here. If this check is done later after the test
314          * of oldbrk with newbrk then it can escape the test and let the data
315          * segment grow beyond its set limit the in case where the limit is
316          * not page aligned -Ram Gupta
317          */
318         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
319                               mm->end_data, mm->start_data))
320                 goto out;
321
322         newbrk = PAGE_ALIGN(brk);
323         oldbrk = PAGE_ALIGN(mm->brk);
324         if (oldbrk == newbrk)
325                 goto set_brk;
326
327         /* Always allow shrinking brk. */
328         if (brk <= mm->brk) {
329                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
330                         goto set_brk;
331                 goto out;
332         }
333
334         /* Check against existing mmap mappings. */
335         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
336                 goto out;
337
338         /* Ok, looks good - let it rip. */
339         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
340                 goto out;
341
342 set_brk:
343         mm->brk = brk;
344         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
345         up_write(&mm->mmap_sem);
346         if (populate)
347                 mm_populate(oldbrk, newbrk - oldbrk);
348         return brk;
349
350 out:
351         retval = mm->brk;
352         up_write(&mm->mmap_sem);
353         return retval;
354 }
355
356 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
357 {
358         unsigned long max, subtree_gap;
359         max = vma->vm_start;
360         if (vma->vm_prev)
361                 max -= vma->vm_prev->vm_end;
362         if (vma->vm_rb.rb_left) {
363                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
364                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
365                 if (subtree_gap > max)
366                         max = subtree_gap;
367         }
368         if (vma->vm_rb.rb_right) {
369                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
370                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
371                 if (subtree_gap > max)
372                         max = subtree_gap;
373         }
374         return max;
375 }
376
377 #ifdef CONFIG_DEBUG_VM_RB
378 static int browse_rb(struct rb_root *root)
379 {
380         int i = 0, j, bug = 0;
381         struct rb_node *nd, *pn = NULL;
382         unsigned long prev = 0, pend = 0;
383
384         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
385                 struct vm_area_struct *vma;
386                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
387                 if (vma->vm_start < prev) {
388                         pr_emerg("vm_start %lx < prev %lx\n",
389                                   vma->vm_start, prev);
390                         bug = 1;
391                 }
392                 if (vma->vm_start < pend) {
393                         pr_emerg("vm_start %lx < pend %lx\n",
394                                   vma->vm_start, pend);
395                         bug = 1;
396                 }
397                 if (vma->vm_start > vma->vm_end) {
398                         pr_emerg("vm_start %lx > vm_end %lx\n",
399                                   vma->vm_start, vma->vm_end);
400                         bug = 1;
401                 }
402                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
403                         pr_emerg("free gap %lx, correct %lx\n",
404                                vma->rb_subtree_gap,
405                                vma_compute_subtree_gap(vma));
406                         bug = 1;
407                 }
408                 i++;
409                 pn = nd;
410                 prev = vma->vm_start;
411                 pend = vma->vm_end;
412         }
413         j = 0;
414         for (nd = pn; nd; nd = rb_prev(nd))
415                 j++;
416         if (i != j) {
417                 pr_emerg("backwards %d, forwards %d\n", j, i);
418                 bug = 1;
419         }
420         return bug ? -1 : i;
421 }
422
423 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
424 {
425         struct rb_node *nd;
426
427         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
428                 struct vm_area_struct *vma;
429                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
430                 VM_BUG_ON_VMA(vma != ignore &&
431                         vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
432                         vma);
433         }
434 }
435
436 static void validate_mm(struct mm_struct *mm)
437 {
438         int bug = 0;
439         int i = 0;
440         unsigned long highest_address = 0;
441         struct vm_area_struct *vma = mm->mmap;
442
443         while (vma) {
444                 struct anon_vma_chain *avc;
445
446                 vma_lock_anon_vma(vma);
447                 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
448                         anon_vma_interval_tree_verify(avc);
449                 vma_unlock_anon_vma(vma);
450                 highest_address = vma->vm_end;
451                 vma = vma->vm_next;
452                 i++;
453         }
454         if (i != mm->map_count) {
455                 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
456                 bug = 1;
457         }
458         if (highest_address != mm->highest_vm_end) {
459                 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
460                           mm->highest_vm_end, highest_address);
461                 bug = 1;
462         }
463         i = browse_rb(&mm->mm_rb);
464         if (i != mm->map_count) {
465                 if (i != -1)
466                         pr_emerg("map_count %d rb %d\n", mm->map_count, i);
467                 bug = 1;
468         }
469         VM_BUG_ON_MM(bug, mm);
470 }
471 #else
472 #define validate_mm_rb(root, ignore) do { } while (0)
473 #define validate_mm(mm) do { } while (0)
474 #endif
475
476 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
477                      unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
478
479 /*
480  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
481  * vma->vm_prev->vm_end values changed, without modifying the vma's position
482  * in the rbtree.
483  */
484 static void vma_gap_update(struct vm_area_struct *vma)
485 {
486         /*
487          * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
488          * function that does exacltly what we want.
489          */
490         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
491 }
492
493 static inline void vma_rb_insert(struct vm_area_struct *vma,
494                                  struct rb_root *root)
495 {
496         /* All rb_subtree_gap values must be consistent prior to insertion */
497         validate_mm_rb(root, NULL);
498
499         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
500 }
501
502 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
503 {
504         /*
505          * All rb_subtree_gap values must be consistent prior to erase,
506          * with the possible exception of the vma being erased.
507          */
508         validate_mm_rb(root, vma);
509
510         /*
511          * Note rb_erase_augmented is a fairly large inline function,
512          * so make sure we instantiate it only once with our desired
513          * augmented rbtree callbacks.
514          */
515         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
516 }
517
518 /*
519  * vma has some anon_vma assigned, and is already inserted on that
520  * anon_vma's interval trees.
521  *
522  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
523  * vma must be removed from the anon_vma's interval trees using
524  * anon_vma_interval_tree_pre_update_vma().
525  *
526  * After the update, the vma will be reinserted using
527  * anon_vma_interval_tree_post_update_vma().
528  *
529  * The entire update must be protected by exclusive mmap_sem and by
530  * the root anon_vma's mutex.
531  */
532 static inline void
533 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
534 {
535         struct anon_vma_chain *avc;
536
537         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
538                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
539 }
540
541 static inline void
542 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
543 {
544         struct anon_vma_chain *avc;
545
546         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
547                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
548 }
549
550 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
551                 unsigned long end, struct vm_area_struct **pprev,
552                 struct rb_node ***rb_link, struct rb_node **rb_parent)
553 {
554         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
555
556         __rb_link = &mm->mm_rb.rb_node;
557         rb_prev = __rb_parent = NULL;
558
559         while (*__rb_link) {
560                 struct vm_area_struct *vma_tmp;
561
562                 __rb_parent = *__rb_link;
563                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
564
565                 if (vma_tmp->vm_end > addr) {
566                         /* Fail if an existing vma overlaps the area */
567                         if (vma_tmp->vm_start < end)
568                                 return -ENOMEM;
569                         __rb_link = &__rb_parent->rb_left;
570                 } else {
571                         rb_prev = __rb_parent;
572                         __rb_link = &__rb_parent->rb_right;
573                 }
574         }
575
576         *pprev = NULL;
577         if (rb_prev)
578                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
579         *rb_link = __rb_link;
580         *rb_parent = __rb_parent;
581         return 0;
582 }
583
584 static unsigned long count_vma_pages_range(struct mm_struct *mm,
585                 unsigned long addr, unsigned long end)
586 {
587         unsigned long nr_pages = 0;
588         struct vm_area_struct *vma;
589
590         /* Find first overlaping mapping */
591         vma = find_vma_intersection(mm, addr, end);
592         if (!vma)
593                 return 0;
594
595         nr_pages = (min(end, vma->vm_end) -
596                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
597
598         /* Iterate over the rest of the overlaps */
599         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
600                 unsigned long overlap_len;
601
602                 if (vma->vm_start > end)
603                         break;
604
605                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
606                 nr_pages += overlap_len >> PAGE_SHIFT;
607         }
608
609         return nr_pages;
610 }
611
612 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
613                 struct rb_node **rb_link, struct rb_node *rb_parent)
614 {
615         /* Update tracking information for the gap following the new vma. */
616         if (vma->vm_next)
617                 vma_gap_update(vma->vm_next);
618         else
619                 mm->highest_vm_end = vma->vm_end;
620
621         /*
622          * vma->vm_prev wasn't known when we followed the rbtree to find the
623          * correct insertion point for that vma. As a result, we could not
624          * update the vma vm_rb parents rb_subtree_gap values on the way down.
625          * So, we first insert the vma with a zero rb_subtree_gap value
626          * (to be consistent with what we did on the way down), and then
627          * immediately update the gap to the correct value. Finally we
628          * rebalance the rbtree after all augmented values have been set.
629          */
630         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
631         vma->rb_subtree_gap = 0;
632         vma_gap_update(vma);
633         vma_rb_insert(vma, &mm->mm_rb);
634 }
635
636 static void __vma_link_file(struct vm_area_struct *vma)
637 {
638         struct file *file;
639
640         file = vma->vm_file;
641         if (file) {
642                 struct address_space *mapping = file->f_mapping;
643
644                 if (vma->vm_flags & VM_DENYWRITE)
645                         atomic_dec(&file_inode(file)->i_writecount);
646                 if (vma->vm_flags & VM_SHARED)
647                         atomic_inc(&mapping->i_mmap_writable);
648
649                 flush_dcache_mmap_lock(mapping);
650                 vma_interval_tree_insert(vma, &mapping->i_mmap);
651                 flush_dcache_mmap_unlock(mapping);
652         }
653 }
654
655 static void
656 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
657         struct vm_area_struct *prev, struct rb_node **rb_link,
658         struct rb_node *rb_parent)
659 {
660         __vma_link_list(mm, vma, prev, rb_parent);
661         __vma_link_rb(mm, vma, rb_link, rb_parent);
662 }
663
664 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
665                         struct vm_area_struct *prev, struct rb_node **rb_link,
666                         struct rb_node *rb_parent)
667 {
668         struct address_space *mapping = NULL;
669
670         if (vma->vm_file) {
671                 mapping = vma->vm_file->f_mapping;
672                 i_mmap_lock_write(mapping);
673         }
674
675         __vma_link(mm, vma, prev, rb_link, rb_parent);
676         __vma_link_file(vma);
677
678         if (mapping)
679                 i_mmap_unlock_write(mapping);
680
681         mm->map_count++;
682         validate_mm(mm);
683 }
684
685 /*
686  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
687  * mm's list and rbtree.  It has already been inserted into the interval tree.
688  */
689 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
690 {
691         struct vm_area_struct *prev;
692         struct rb_node **rb_link, *rb_parent;
693
694         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
695                            &prev, &rb_link, &rb_parent))
696                 BUG();
697         __vma_link(mm, vma, prev, rb_link, rb_parent);
698         mm->map_count++;
699 }
700
701 static inline void
702 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
703                 struct vm_area_struct *prev)
704 {
705         struct vm_area_struct *next;
706
707         vma_rb_erase(vma, &mm->mm_rb);
708         prev->vm_next = next = vma->vm_next;
709         if (next)
710                 next->vm_prev = prev;
711
712         /* Kill the cache */
713         vmacache_invalidate(mm);
714 }
715
716 /*
717  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
718  * is already present in an i_mmap tree without adjusting the tree.
719  * The following helper function should be used when such adjustments
720  * are necessary.  The "insert" vma (if any) is to be inserted
721  * before we drop the necessary locks.
722  */
723 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
724         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
725 {
726         struct mm_struct *mm = vma->vm_mm;
727         struct vm_area_struct *next = vma->vm_next;
728         struct vm_area_struct *importer = NULL;
729         struct address_space *mapping = NULL;
730         struct rb_root *root = NULL;
731         struct anon_vma *anon_vma = NULL;
732         struct file *file = vma->vm_file;
733         bool start_changed = false, end_changed = false;
734         long adjust_next = 0;
735         int remove_next = 0;
736
737         if (next && !insert) {
738                 struct vm_area_struct *exporter = NULL;
739
740                 if (end >= next->vm_end) {
741                         /*
742                          * vma expands, overlapping all the next, and
743                          * perhaps the one after too (mprotect case 6).
744                          */
745 again:                  remove_next = 1 + (end > next->vm_end);
746                         end = next->vm_end;
747                         exporter = next;
748                         importer = vma;
749                 } else if (end > next->vm_start) {
750                         /*
751                          * vma expands, overlapping part of the next:
752                          * mprotect case 5 shifting the boundary up.
753                          */
754                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
755                         exporter = next;
756                         importer = vma;
757                 } else if (end < vma->vm_end) {
758                         /*
759                          * vma shrinks, and !insert tells it's not
760                          * split_vma inserting another: so it must be
761                          * mprotect case 4 shifting the boundary down.
762                          */
763                         adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
764                         exporter = vma;
765                         importer = next;
766                 }
767
768                 /*
769                  * Easily overlooked: when mprotect shifts the boundary,
770                  * make sure the expanding vma has anon_vma set if the
771                  * shrinking vma had, to cover any anon pages imported.
772                  */
773                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
774                         int error;
775
776                         importer->anon_vma = exporter->anon_vma;
777                         error = anon_vma_clone(importer, exporter);
778                         if (error)
779                                 return error;
780                 }
781         }
782
783         if (file) {
784                 mapping = file->f_mapping;
785                 root = &mapping->i_mmap;
786                 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
787
788                 if (adjust_next)
789                         uprobe_munmap(next, next->vm_start, next->vm_end);
790
791                 i_mmap_lock_write(mapping);
792                 if (insert) {
793                         /*
794                          * Put into interval tree now, so instantiated pages
795                          * are visible to arm/parisc __flush_dcache_page
796                          * throughout; but we cannot insert into address
797                          * space until vma start or end is updated.
798                          */
799                         __vma_link_file(insert);
800                 }
801         }
802
803         vma_adjust_trans_huge(vma, start, end, adjust_next);
804
805         anon_vma = vma->anon_vma;
806         if (!anon_vma && adjust_next)
807                 anon_vma = next->anon_vma;
808         if (anon_vma) {
809                 VM_BUG_ON_VMA(adjust_next && next->anon_vma &&
810                           anon_vma != next->anon_vma, next);
811                 anon_vma_lock_write(anon_vma);
812                 anon_vma_interval_tree_pre_update_vma(vma);
813                 if (adjust_next)
814                         anon_vma_interval_tree_pre_update_vma(next);
815         }
816
817         if (root) {
818                 flush_dcache_mmap_lock(mapping);
819                 vma_interval_tree_remove(vma, root);
820                 if (adjust_next)
821                         vma_interval_tree_remove(next, root);
822         }
823
824         if (start != vma->vm_start) {
825                 vma->vm_start = start;
826                 start_changed = true;
827         }
828         if (end != vma->vm_end) {
829                 vma->vm_end = end;
830                 end_changed = true;
831         }
832         vma->vm_pgoff = pgoff;
833         if (adjust_next) {
834                 next->vm_start += adjust_next << PAGE_SHIFT;
835                 next->vm_pgoff += adjust_next;
836         }
837
838         if (root) {
839                 if (adjust_next)
840                         vma_interval_tree_insert(next, root);
841                 vma_interval_tree_insert(vma, root);
842                 flush_dcache_mmap_unlock(mapping);
843         }
844
845         if (remove_next) {
846                 /*
847                  * vma_merge has merged next into vma, and needs
848                  * us to remove next before dropping the locks.
849                  */
850                 __vma_unlink(mm, next, vma);
851                 if (file)
852                         __remove_shared_vm_struct(next, file, mapping);
853         } else if (insert) {
854                 /*
855                  * split_vma has split insert from vma, and needs
856                  * us to insert it before dropping the locks
857                  * (it may either follow vma or precede it).
858                  */
859                 __insert_vm_struct(mm, insert);
860         } else {
861                 if (start_changed)
862                         vma_gap_update(vma);
863                 if (end_changed) {
864                         if (!next)
865                                 mm->highest_vm_end = end;
866                         else if (!adjust_next)
867                                 vma_gap_update(next);
868                 }
869         }
870
871         if (anon_vma) {
872                 anon_vma_interval_tree_post_update_vma(vma);
873                 if (adjust_next)
874                         anon_vma_interval_tree_post_update_vma(next);
875                 anon_vma_unlock_write(anon_vma);
876         }
877         if (mapping)
878                 i_mmap_unlock_write(mapping);
879
880         if (root) {
881                 uprobe_mmap(vma);
882
883                 if (adjust_next)
884                         uprobe_mmap(next);
885         }
886
887         if (remove_next) {
888                 if (file) {
889                         uprobe_munmap(next, next->vm_start, next->vm_end);
890                         fput(file);
891                 }
892                 if (next->anon_vma)
893                         anon_vma_merge(vma, next);
894                 mm->map_count--;
895                 mpol_put(vma_policy(next));
896                 kmem_cache_free(vm_area_cachep, next);
897                 /*
898                  * In mprotect's case 6 (see comments on vma_merge),
899                  * we must remove another next too. It would clutter
900                  * up the code too much to do both in one go.
901                  */
902                 next = vma->vm_next;
903                 if (remove_next == 2)
904                         goto again;
905                 else if (next)
906                         vma_gap_update(next);
907                 else
908                         mm->highest_vm_end = end;
909         }
910         if (insert && file)
911                 uprobe_mmap(insert);
912
913         validate_mm(mm);
914
915         return 0;
916 }
917
918 /*
919  * If the vma has a ->close operation then the driver probably needs to release
920  * per-vma resources, so we don't attempt to merge those.
921  */
922 static inline int is_mergeable_vma(struct vm_area_struct *vma,
923                                 struct file *file, unsigned long vm_flags,
924                                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
925 {
926         /*
927          * VM_SOFTDIRTY should not prevent from VMA merging, if we
928          * match the flags but dirty bit -- the caller should mark
929          * merged VMA as dirty. If dirty bit won't be excluded from
930          * comparison, we increase pressue on the memory system forcing
931          * the kernel to generate new VMAs when old one could be
932          * extended instead.
933          */
934         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
935                 return 0;
936         if (vma->vm_file != file)
937                 return 0;
938         if (vma->vm_ops && vma->vm_ops->close)
939                 return 0;
940         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
941                 return 0;
942         return 1;
943 }
944
945 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
946                                         struct anon_vma *anon_vma2,
947                                         struct vm_area_struct *vma)
948 {
949         /*
950          * The list_is_singular() test is to avoid merging VMA cloned from
951          * parents. This can improve scalability caused by anon_vma lock.
952          */
953         if ((!anon_vma1 || !anon_vma2) && (!vma ||
954                 list_is_singular(&vma->anon_vma_chain)))
955                 return 1;
956         return anon_vma1 == anon_vma2;
957 }
958
959 /*
960  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
961  * in front of (at a lower virtual address and file offset than) the vma.
962  *
963  * We cannot merge two vmas if they have differently assigned (non-NULL)
964  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
965  *
966  * We don't check here for the merged mmap wrapping around the end of pagecache
967  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
968  * wrap, nor mmaps which cover the final page at index -1UL.
969  */
970 static int
971 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
972                      struct anon_vma *anon_vma, struct file *file,
973                      pgoff_t vm_pgoff,
974                      struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
975 {
976         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
977             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
978                 if (vma->vm_pgoff == vm_pgoff)
979                         return 1;
980         }
981         return 0;
982 }
983
984 /*
985  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
986  * beyond (at a higher virtual address and file offset than) the vma.
987  *
988  * We cannot merge two vmas if they have differently assigned (non-NULL)
989  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
990  */
991 static int
992 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
993                     struct anon_vma *anon_vma, struct file *file,
994                     pgoff_t vm_pgoff,
995                     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
996 {
997         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
998             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
999                 pgoff_t vm_pglen;
1000                 vm_pglen = vma_pages(vma);
1001                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1002                         return 1;
1003         }
1004         return 0;
1005 }
1006
1007 /*
1008  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1009  * whether that can be merged with its predecessor or its successor.
1010  * Or both (it neatly fills a hole).
1011  *
1012  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1013  * certain not to be mapped by the time vma_merge is called; but when
1014  * called for mprotect, it is certain to be already mapped (either at
1015  * an offset within prev, or at the start of next), and the flags of
1016  * this area are about to be changed to vm_flags - and the no-change
1017  * case has already been eliminated.
1018  *
1019  * The following mprotect cases have to be considered, where AAAA is
1020  * the area passed down from mprotect_fixup, never extending beyond one
1021  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1022  *
1023  *     AAAA             AAAA                AAAA          AAAA
1024  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
1025  *    cannot merge    might become    might become    might become
1026  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
1027  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
1028  *    mremap move:                                    PPPPNNNNNNNN 8
1029  *        AAAA
1030  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
1031  *    might become    case 1 below    case 2 below    case 3 below
1032  *
1033  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1034  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1035  */
1036 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1037                         struct vm_area_struct *prev, unsigned long addr,
1038                         unsigned long end, unsigned long vm_flags,
1039                         struct anon_vma *anon_vma, struct file *file,
1040                         pgoff_t pgoff, struct mempolicy *policy,
1041                         struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1042 {
1043         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1044         struct vm_area_struct *area, *next;
1045         int err;
1046
1047         /*
1048          * We later require that vma->vm_flags == vm_flags,
1049          * so this tests vma->vm_flags & VM_SPECIAL, too.
1050          */
1051         if (vm_flags & VM_SPECIAL)
1052                 return NULL;
1053
1054         if (prev)
1055                 next = prev->vm_next;
1056         else
1057                 next = mm->mmap;
1058         area = next;
1059         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
1060                 next = next->vm_next;
1061
1062         /*
1063          * Can it merge with the predecessor?
1064          */
1065         if (prev && prev->vm_end == addr &&
1066                         mpol_equal(vma_policy(prev), policy) &&
1067                         can_vma_merge_after(prev, vm_flags,
1068                                             anon_vma, file, pgoff,
1069                                             vm_userfaultfd_ctx)) {
1070                 /*
1071                  * OK, it can.  Can we now merge in the successor as well?
1072                  */
1073                 if (next && end == next->vm_start &&
1074                                 mpol_equal(policy, vma_policy(next)) &&
1075                                 can_vma_merge_before(next, vm_flags,
1076                                                      anon_vma, file,
1077                                                      pgoff+pglen,
1078                                                      vm_userfaultfd_ctx) &&
1079                                 is_mergeable_anon_vma(prev->anon_vma,
1080                                                       next->anon_vma, NULL)) {
1081                                                         /* cases 1, 6 */
1082                         err = vma_adjust(prev, prev->vm_start,
1083                                 next->vm_end, prev->vm_pgoff, NULL);
1084                 } else                                  /* cases 2, 5, 7 */
1085                         err = vma_adjust(prev, prev->vm_start,
1086                                 end, prev->vm_pgoff, NULL);
1087                 if (err)
1088                         return NULL;
1089                 khugepaged_enter_vma_merge(prev, vm_flags);
1090                 return prev;
1091         }
1092
1093         /*
1094          * Can this new request be merged in front of next?
1095          */
1096         if (next && end == next->vm_start &&
1097                         mpol_equal(policy, vma_policy(next)) &&
1098                         can_vma_merge_before(next, vm_flags,
1099                                              anon_vma, file, pgoff+pglen,
1100                                              vm_userfaultfd_ctx)) {
1101                 if (prev && addr < prev->vm_end)        /* case 4 */
1102                         err = vma_adjust(prev, prev->vm_start,
1103                                 addr, prev->vm_pgoff, NULL);
1104                 else                                    /* cases 3, 8 */
1105                         err = vma_adjust(area, addr, next->vm_end,
1106                                 next->vm_pgoff - pglen, NULL);
1107                 if (err)
1108                         return NULL;
1109                 khugepaged_enter_vma_merge(area, vm_flags);
1110                 return area;
1111         }
1112
1113         return NULL;
1114 }
1115
1116 /*
1117  * Rough compatbility check to quickly see if it's even worth looking
1118  * at sharing an anon_vma.
1119  *
1120  * They need to have the same vm_file, and the flags can only differ
1121  * in things that mprotect may change.
1122  *
1123  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1124  * we can merge the two vma's. For example, we refuse to merge a vma if
1125  * there is a vm_ops->close() function, because that indicates that the
1126  * driver is doing some kind of reference counting. But that doesn't
1127  * really matter for the anon_vma sharing case.
1128  */
1129 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1130 {
1131         return a->vm_end == b->vm_start &&
1132                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1133                 a->vm_file == b->vm_file &&
1134                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1135                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1136 }
1137
1138 /*
1139  * Do some basic sanity checking to see if we can re-use the anon_vma
1140  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1141  * the same as 'old', the other will be the new one that is trying
1142  * to share the anon_vma.
1143  *
1144  * NOTE! This runs with mm_sem held for reading, so it is possible that
1145  * the anon_vma of 'old' is concurrently in the process of being set up
1146  * by another page fault trying to merge _that_. But that's ok: if it
1147  * is being set up, that automatically means that it will be a singleton
1148  * acceptable for merging, so we can do all of this optimistically. But
1149  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1150  *
1151  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1152  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1153  * is to return an anon_vma that is "complex" due to having gone through
1154  * a fork).
1155  *
1156  * We also make sure that the two vma's are compatible (adjacent,
1157  * and with the same memory policies). That's all stable, even with just
1158  * a read lock on the mm_sem.
1159  */
1160 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1161 {
1162         if (anon_vma_compatible(a, b)) {
1163                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1164
1165                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1166                         return anon_vma;
1167         }
1168         return NULL;
1169 }
1170
1171 /*
1172  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1173  * neighbouring vmas for a suitable anon_vma, before it goes off
1174  * to allocate a new anon_vma.  It checks because a repetitive
1175  * sequence of mprotects and faults may otherwise lead to distinct
1176  * anon_vmas being allocated, preventing vma merge in subsequent
1177  * mprotect.
1178  */
1179 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1180 {
1181         struct anon_vma *anon_vma;
1182         struct vm_area_struct *near;
1183
1184         near = vma->vm_next;
1185         if (!near)
1186                 goto try_prev;
1187
1188         anon_vma = reusable_anon_vma(near, vma, near);
1189         if (anon_vma)
1190                 return anon_vma;
1191 try_prev:
1192         near = vma->vm_prev;
1193         if (!near)
1194                 goto none;
1195
1196         anon_vma = reusable_anon_vma(near, near, vma);
1197         if (anon_vma)
1198                 return anon_vma;
1199 none:
1200         /*
1201          * There's no absolute need to look only at touching neighbours:
1202          * we could search further afield for "compatible" anon_vmas.
1203          * But it would probably just be a waste of time searching,
1204          * or lead to too many vmas hanging off the same anon_vma.
1205          * We're trying to allow mprotect remerging later on,
1206          * not trying to minimize memory used for anon_vmas.
1207          */
1208         return NULL;
1209 }
1210
1211 #ifdef CONFIG_PROC_FS
1212 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1213                                                 struct file *file, long pages)
1214 {
1215         const unsigned long stack_flags
1216                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1217
1218         mm->total_vm += pages;
1219
1220         if (file) {
1221                 mm->shared_vm += pages;
1222                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1223                         mm->exec_vm += pages;
1224         } else if (flags & stack_flags)
1225                 mm->stack_vm += pages;
1226 }
1227 #endif /* CONFIG_PROC_FS */
1228
1229 /*
1230  * If a hint addr is less than mmap_min_addr change hint to be as
1231  * low as possible but still greater than mmap_min_addr
1232  */
1233 static inline unsigned long round_hint_to_min(unsigned long hint)
1234 {
1235         hint &= PAGE_MASK;
1236         if (((void *)hint != NULL) &&
1237             (hint < mmap_min_addr))
1238                 return PAGE_ALIGN(mmap_min_addr);
1239         return hint;
1240 }
1241
1242 static inline int mlock_future_check(struct mm_struct *mm,
1243                                      unsigned long flags,
1244                                      unsigned long len)
1245 {
1246         unsigned long locked, lock_limit;
1247
1248         /*  mlock MCL_FUTURE? */
1249         if (flags & VM_LOCKED) {
1250                 locked = len >> PAGE_SHIFT;
1251                 locked += mm->locked_vm;
1252                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1253                 lock_limit >>= PAGE_SHIFT;
1254                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1255                         return -EAGAIN;
1256         }
1257         return 0;
1258 }
1259
1260 /*
1261  * The caller must hold down_write(&current->mm->mmap_sem).
1262  */
1263 unsigned long do_mmap(struct file *file, unsigned long addr,
1264                         unsigned long len, unsigned long prot,
1265                         unsigned long flags, vm_flags_t vm_flags,
1266                         unsigned long pgoff, unsigned long *populate)
1267 {
1268         struct mm_struct *mm = current->mm;
1269
1270         *populate = 0;
1271
1272         if (!len)
1273                 return -EINVAL;
1274
1275         /*
1276          * Does the application expect PROT_READ to imply PROT_EXEC?
1277          *
1278          * (the exception is when the underlying filesystem is noexec
1279          *  mounted, in which case we dont add PROT_EXEC.)
1280          */
1281         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1282                 if (!(file && path_noexec(&file->f_path)))
1283                         prot |= PROT_EXEC;
1284
1285         if (!(flags & MAP_FIXED))
1286                 addr = round_hint_to_min(addr);
1287
1288         /* Careful about overflows.. */
1289         len = PAGE_ALIGN(len);
1290         if (!len)
1291                 return -ENOMEM;
1292
1293         /* offset overflow? */
1294         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1295                 return -EOVERFLOW;
1296
1297         /* Too many mappings? */
1298         if (mm->map_count > sysctl_max_map_count)
1299                 return -ENOMEM;
1300
1301         /* Obtain the address to map to. we verify (or select) it and ensure
1302          * that it represents a valid section of the address space.
1303          */
1304         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1305         if (addr & ~PAGE_MASK)
1306                 return addr;
1307
1308         /* Do simple checking here so the lower-level routines won't have
1309          * to. we assume access permissions have been handled by the open
1310          * of the memory object, so we don't do any here.
1311          */
1312         vm_flags |= calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1313                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1314
1315         if (flags & MAP_LOCKED)
1316                 if (!can_do_mlock())
1317                         return -EPERM;
1318
1319         if (mlock_future_check(mm, vm_flags, len))
1320                 return -EAGAIN;
1321
1322         if (file) {
1323                 struct inode *inode = file_inode(file);
1324
1325                 switch (flags & MAP_TYPE) {
1326                 case MAP_SHARED:
1327                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1328                                 return -EACCES;
1329
1330                         /*
1331                          * Make sure we don't allow writing to an append-only
1332                          * file..
1333                          */
1334                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1335                                 return -EACCES;
1336
1337                         /*
1338                          * Make sure there are no mandatory locks on the file.
1339                          */
1340                         if (locks_verify_locked(file))
1341                                 return -EAGAIN;
1342
1343                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1344                         if (!(file->f_mode & FMODE_WRITE))
1345                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1346
1347                         /* fall through */
1348                 case MAP_PRIVATE:
1349                         if (!(file->f_mode & FMODE_READ))
1350                                 return -EACCES;
1351                         if (path_noexec(&file->f_path)) {
1352                                 if (vm_flags & VM_EXEC)
1353                                         return -EPERM;
1354                                 vm_flags &= ~VM_MAYEXEC;
1355                         }
1356
1357                         if (!file->f_op->mmap)
1358                                 return -ENODEV;
1359                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1360                                 return -EINVAL;
1361                         break;
1362
1363                 default:
1364                         return -EINVAL;
1365                 }
1366         } else {
1367                 switch (flags & MAP_TYPE) {
1368                 case MAP_SHARED:
1369                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1370                                 return -EINVAL;
1371                         /*
1372                          * Ignore pgoff.
1373                          */
1374                         pgoff = 0;
1375                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1376                         break;
1377                 case MAP_PRIVATE:
1378                         /*
1379                          * Set pgoff according to addr for anon_vma.
1380                          */
1381                         pgoff = addr >> PAGE_SHIFT;
1382                         break;
1383                 default:
1384                         return -EINVAL;
1385                 }
1386         }
1387
1388         /*
1389          * Set 'VM_NORESERVE' if we should not account for the
1390          * memory use of this mapping.
1391          */
1392         if (flags & MAP_NORESERVE) {
1393                 /* We honor MAP_NORESERVE if allowed to overcommit */
1394                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1395                         vm_flags |= VM_NORESERVE;
1396
1397                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1398                 if (file && is_file_hugepages(file))
1399                         vm_flags |= VM_NORESERVE;
1400         }
1401
1402         addr = mmap_region(file, addr, len, vm_flags, pgoff);
1403         if (!IS_ERR_VALUE(addr) &&
1404             ((vm_flags & VM_LOCKED) ||
1405              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1406                 *populate = len;
1407         return addr;
1408 }
1409
1410 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1411                 unsigned long, prot, unsigned long, flags,
1412                 unsigned long, fd, unsigned long, pgoff)
1413 {
1414         struct file *file = NULL;
1415         unsigned long retval = -EBADF;
1416
1417         if (!(flags & MAP_ANONYMOUS)) {
1418                 audit_mmap_fd(fd, flags);
1419                 file = fget(fd);
1420                 if (!file)
1421                         goto out;
1422                 if (is_file_hugepages(file))
1423                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1424                 retval = -EINVAL;
1425                 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1426                         goto out_fput;
1427         } else if (flags & MAP_HUGETLB) {
1428                 struct user_struct *user = NULL;
1429                 struct hstate *hs;
1430
1431                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1432                 if (!hs)
1433                         return -EINVAL;
1434
1435                 len = ALIGN(len, huge_page_size(hs));
1436                 /*
1437                  * VM_NORESERVE is used because the reservations will be
1438                  * taken when vm_ops->mmap() is called
1439                  * A dummy user value is used because we are not locking
1440                  * memory so no accounting is necessary
1441                  */
1442                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1443                                 VM_NORESERVE,
1444                                 &user, HUGETLB_ANONHUGE_INODE,
1445                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1446                 if (IS_ERR(file))
1447                         return PTR_ERR(file);
1448         }
1449
1450         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1451
1452         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1453 out_fput:
1454         if (file)
1455                 fput(file);
1456 out:
1457         return retval;
1458 }
1459
1460 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1461 struct mmap_arg_struct {
1462         unsigned long addr;
1463         unsigned long len;
1464         unsigned long prot;
1465         unsigned long flags;
1466         unsigned long fd;
1467         unsigned long offset;
1468 };
1469
1470 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1471 {
1472         struct mmap_arg_struct a;
1473
1474         if (copy_from_user(&a, arg, sizeof(a)))
1475                 return -EFAULT;
1476         if (a.offset & ~PAGE_MASK)
1477                 return -EINVAL;
1478
1479         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1480                               a.offset >> PAGE_SHIFT);
1481 }
1482 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1483
1484 /*
1485  * Some shared mappigns will want the pages marked read-only
1486  * to track write events. If so, we'll downgrade vm_page_prot
1487  * to the private version (using protection_map[] without the
1488  * VM_SHARED bit).
1489  */
1490 int vma_wants_writenotify(struct vm_area_struct *vma)
1491 {
1492         vm_flags_t vm_flags = vma->vm_flags;
1493
1494         /* If it was private or non-writable, the write bit is already clear */
1495         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1496                 return 0;
1497
1498         /* The backer wishes to know when pages are first written to? */
1499         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1500                 return 1;
1501
1502         /* The open routine did something to the protections that pgprot_modify
1503          * won't preserve? */
1504         if (pgprot_val(vma->vm_page_prot) !=
1505             pgprot_val(vm_pgprot_modify(vma->vm_page_prot, vm_flags)))
1506                 return 0;
1507
1508         /* Do we need to track softdirty? */
1509         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1510                 return 1;
1511
1512         /* Specialty mapping? */
1513         if (vm_flags & VM_PFNMAP)
1514                 return 0;
1515
1516         /* Can the mapping track the dirty pages? */
1517         return vma->vm_file && vma->vm_file->f_mapping &&
1518                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1519 }
1520
1521 /*
1522  * We account for memory if it's a private writeable mapping,
1523  * not hugepages and VM_NORESERVE wasn't set.
1524  */
1525 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1526 {
1527         /*
1528          * hugetlb has its own accounting separate from the core VM
1529          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1530          */
1531         if (file && is_file_hugepages(file))
1532                 return 0;
1533
1534         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1535 }
1536
1537 unsigned long mmap_region(struct file *file, unsigned long addr,
1538                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1539 {
1540         struct mm_struct *mm = current->mm;
1541         struct vm_area_struct *vma, *prev;
1542         int error;
1543         struct rb_node **rb_link, *rb_parent;
1544         unsigned long charged = 0;
1545
1546         /* Check against address space limit. */
1547         if (!may_expand_vm(mm, len >> PAGE_SHIFT)) {
1548                 unsigned long nr_pages;
1549
1550                 /*
1551                  * MAP_FIXED may remove pages of mappings that intersects with
1552                  * requested mapping. Account for the pages it would unmap.
1553                  */
1554                 if (!(vm_flags & MAP_FIXED))
1555                         return -ENOMEM;
1556
1557                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1558
1559                 if (!may_expand_vm(mm, (len >> PAGE_SHIFT) - nr_pages))
1560                         return -ENOMEM;
1561         }
1562
1563         /* Clear old maps */
1564         error = -ENOMEM;
1565         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1566                               &rb_parent)) {
1567                 if (do_munmap(mm, addr, len))
1568                         return -ENOMEM;
1569         }
1570
1571         /*
1572          * Private writable mapping: check memory availability
1573          */
1574         if (accountable_mapping(file, vm_flags)) {
1575                 charged = len >> PAGE_SHIFT;
1576                 if (security_vm_enough_memory_mm(mm, charged))
1577                         return -ENOMEM;
1578                 vm_flags |= VM_ACCOUNT;
1579         }
1580
1581         /*
1582          * Can we just expand an old mapping?
1583          */
1584         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1585                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1586         if (vma)
1587                 goto out;
1588
1589         /*
1590          * Determine the object being mapped and call the appropriate
1591          * specific mapper. the address has already been validated, but
1592          * not unmapped, but the maps are removed from the list.
1593          */
1594         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1595         if (!vma) {
1596                 error = -ENOMEM;
1597                 goto unacct_error;
1598         }
1599
1600         vma->vm_mm = mm;
1601         vma->vm_start = addr;
1602         vma->vm_end = addr + len;
1603         vma->vm_flags = vm_flags;
1604         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1605         vma->vm_pgoff = pgoff;
1606         INIT_LIST_HEAD(&vma->anon_vma_chain);
1607
1608         if (file) {
1609                 if (vm_flags & VM_DENYWRITE) {
1610                         error = deny_write_access(file);
1611                         if (error)
1612                                 goto free_vma;
1613                 }
1614                 if (vm_flags & VM_SHARED) {
1615                         error = mapping_map_writable(file->f_mapping);
1616                         if (error)
1617                                 goto allow_write_and_free_vma;
1618                 }
1619
1620                 /* ->mmap() can change vma->vm_file, but must guarantee that
1621                  * vma_link() below can deny write-access if VM_DENYWRITE is set
1622                  * and map writably if VM_SHARED is set. This usually means the
1623                  * new file must not have been exposed to user-space, yet.
1624                  */
1625                 vma->vm_file = get_file(file);
1626                 error = file->f_op->mmap(file, vma);
1627                 if (error)
1628                         goto unmap_and_free_vma;
1629
1630                 /* Can addr have changed??
1631                  *
1632                  * Answer: Yes, several device drivers can do it in their
1633                  *         f_op->mmap method. -DaveM
1634                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1635                  *      be updated for vma_link()
1636                  */
1637                 WARN_ON_ONCE(addr != vma->vm_start);
1638
1639                 addr = vma->vm_start;
1640                 vm_flags = vma->vm_flags;
1641         } else if (vm_flags & VM_SHARED) {
1642                 error = shmem_zero_setup(vma);
1643                 if (error)
1644                         goto free_vma;
1645         }
1646
1647         vma_link(mm, vma, prev, rb_link, rb_parent);
1648         /* Once vma denies write, undo our temporary denial count */
1649         if (file) {
1650                 if (vm_flags & VM_SHARED)
1651                         mapping_unmap_writable(file->f_mapping);
1652                 if (vm_flags & VM_DENYWRITE)
1653                         allow_write_access(file);
1654         }
1655         file = vma->vm_file;
1656 out:
1657         perf_event_mmap(vma);
1658
1659         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1660         if (vm_flags & VM_LOCKED) {
1661                 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1662                                         vma == get_gate_vma(current->mm)))
1663                         mm->locked_vm += (len >> PAGE_SHIFT);
1664                 else
1665                         vma->vm_flags &= ~VM_LOCKED;
1666         }
1667
1668         if (file)
1669                 uprobe_mmap(vma);
1670
1671         /*
1672          * New (or expanded) vma always get soft dirty status.
1673          * Otherwise user-space soft-dirty page tracker won't
1674          * be able to distinguish situation when vma area unmapped,
1675          * then new mapped in-place (which must be aimed as
1676          * a completely new data area).
1677          */
1678         vma->vm_flags |= VM_SOFTDIRTY;
1679
1680         vma_set_page_prot(vma);
1681
1682         return addr;
1683
1684 unmap_and_free_vma:
1685         vma->vm_file = NULL;
1686         fput(file);
1687
1688         /* Undo any partial mapping done by a device driver. */
1689         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1690         charged = 0;
1691         if (vm_flags & VM_SHARED)
1692                 mapping_unmap_writable(file->f_mapping);
1693 allow_write_and_free_vma:
1694         if (vm_flags & VM_DENYWRITE)
1695                 allow_write_access(file);
1696 free_vma:
1697         kmem_cache_free(vm_area_cachep, vma);
1698 unacct_error:
1699         if (charged)
1700                 vm_unacct_memory(charged);
1701         return error;
1702 }
1703
1704 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1705 {
1706         /*
1707          * We implement the search by looking for an rbtree node that
1708          * immediately follows a suitable gap. That is,
1709          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1710          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1711          * - gap_end - gap_start >= length
1712          */
1713
1714         struct mm_struct *mm = current->mm;
1715         struct vm_area_struct *vma;
1716         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1717
1718         /* Adjust search length to account for worst case alignment overhead */
1719         length = info->length + info->align_mask;
1720         if (length < info->length)
1721                 return -ENOMEM;
1722
1723         /* Adjust search limits by the desired length */
1724         if (info->high_limit < length)
1725                 return -ENOMEM;
1726         high_limit = info->high_limit - length;
1727
1728         if (info->low_limit > high_limit)
1729                 return -ENOMEM;
1730         low_limit = info->low_limit + length;
1731
1732         /* Check if rbtree root looks promising */
1733         if (RB_EMPTY_ROOT(&mm->mm_rb))
1734                 goto check_highest;
1735         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1736         if (vma->rb_subtree_gap < length)
1737                 goto check_highest;
1738
1739         while (true) {
1740                 /* Visit left subtree if it looks promising */
1741                 gap_end = vma->vm_start;
1742                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1743                         struct vm_area_struct *left =
1744                                 rb_entry(vma->vm_rb.rb_left,
1745                                          struct vm_area_struct, vm_rb);
1746                         if (left->rb_subtree_gap >= length) {
1747                                 vma = left;
1748                                 continue;
1749                         }
1750                 }
1751
1752                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1753 check_current:
1754                 /* Check if current node has a suitable gap */
1755                 if (gap_start > high_limit)
1756                         return -ENOMEM;
1757                 if (gap_end >= low_limit && gap_end - gap_start >= length)
1758                         goto found;
1759
1760                 /* Visit right subtree if it looks promising */
1761                 if (vma->vm_rb.rb_right) {
1762                         struct vm_area_struct *right =
1763                                 rb_entry(vma->vm_rb.rb_right,
1764                                          struct vm_area_struct, vm_rb);
1765                         if (right->rb_subtree_gap >= length) {
1766                                 vma = right;
1767                                 continue;
1768                         }
1769                 }
1770
1771                 /* Go back up the rbtree to find next candidate node */
1772                 while (true) {
1773                         struct rb_node *prev = &vma->vm_rb;
1774                         if (!rb_parent(prev))
1775                                 goto check_highest;
1776                         vma = rb_entry(rb_parent(prev),
1777                                        struct vm_area_struct, vm_rb);
1778                         if (prev == vma->vm_rb.rb_left) {
1779                                 gap_start = vma->vm_prev->vm_end;
1780                                 gap_end = vma->vm_start;
1781                                 goto check_current;
1782                         }
1783                 }
1784         }
1785
1786 check_highest:
1787         /* Check highest gap, which does not precede any rbtree node */
1788         gap_start = mm->highest_vm_end;
1789         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1790         if (gap_start > high_limit)
1791                 return -ENOMEM;
1792
1793 found:
1794         /* We found a suitable gap. Clip it with the original low_limit. */
1795         if (gap_start < info->low_limit)
1796                 gap_start = info->low_limit;
1797
1798         /* Adjust gap address to the desired alignment */
1799         gap_start += (info->align_offset - gap_start) & info->align_mask;
1800
1801         VM_BUG_ON(gap_start + info->length > info->high_limit);
1802         VM_BUG_ON(gap_start + info->length > gap_end);
1803         return gap_start;
1804 }
1805
1806 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1807 {
1808         struct mm_struct *mm = current->mm;
1809         struct vm_area_struct *vma;
1810         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1811
1812         /* Adjust search length to account for worst case alignment overhead */
1813         length = info->length + info->align_mask;
1814         if (length < info->length)
1815                 return -ENOMEM;
1816
1817         /*
1818          * Adjust search limits by the desired length.
1819          * See implementation comment at top of unmapped_area().
1820          */
1821         gap_end = info->high_limit;
1822         if (gap_end < length)
1823                 return -ENOMEM;
1824         high_limit = gap_end - length;
1825
1826         if (info->low_limit > high_limit)
1827                 return -ENOMEM;
1828         low_limit = info->low_limit + length;
1829
1830         /* Check highest gap, which does not precede any rbtree node */
1831         gap_start = mm->highest_vm_end;
1832         if (gap_start <= high_limit)
1833                 goto found_highest;
1834
1835         /* Check if rbtree root looks promising */
1836         if (RB_EMPTY_ROOT(&mm->mm_rb))
1837                 return -ENOMEM;
1838         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1839         if (vma->rb_subtree_gap < length)
1840                 return -ENOMEM;
1841
1842         while (true) {
1843                 /* Visit right subtree if it looks promising */
1844                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1845                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1846                         struct vm_area_struct *right =
1847                                 rb_entry(vma->vm_rb.rb_right,
1848                                          struct vm_area_struct, vm_rb);
1849                         if (right->rb_subtree_gap >= length) {
1850                                 vma = right;
1851                                 continue;
1852                         }
1853                 }
1854
1855 check_current:
1856                 /* Check if current node has a suitable gap */
1857                 gap_end = vma->vm_start;
1858                 if (gap_end < low_limit)
1859                         return -ENOMEM;
1860                 if (gap_start <= high_limit && gap_end - gap_start >= length)
1861                         goto found;
1862
1863                 /* Visit left subtree if it looks promising */
1864                 if (vma->vm_rb.rb_left) {
1865                         struct vm_area_struct *left =
1866                                 rb_entry(vma->vm_rb.rb_left,
1867                                          struct vm_area_struct, vm_rb);
1868                         if (left->rb_subtree_gap >= length) {
1869                                 vma = left;
1870                                 continue;
1871                         }
1872                 }
1873
1874                 /* Go back up the rbtree to find next candidate node */
1875                 while (true) {
1876                         struct rb_node *prev = &vma->vm_rb;
1877                         if (!rb_parent(prev))
1878                                 return -ENOMEM;
1879                         vma = rb_entry(rb_parent(prev),
1880                                        struct vm_area_struct, vm_rb);
1881                         if (prev == vma->vm_rb.rb_right) {
1882                                 gap_start = vma->vm_prev ?
1883                                         vma->vm_prev->vm_end : 0;
1884                                 goto check_current;
1885                         }
1886                 }
1887         }
1888
1889 found:
1890         /* We found a suitable gap. Clip it with the original high_limit. */
1891         if (gap_end > info->high_limit)
1892                 gap_end = info->high_limit;
1893
1894 found_highest:
1895         /* Compute highest gap address at the desired alignment */
1896         gap_end -= info->length;
1897         gap_end -= (gap_end - info->align_offset) & info->align_mask;
1898
1899         VM_BUG_ON(gap_end < info->low_limit);
1900         VM_BUG_ON(gap_end < gap_start);
1901         return gap_end;
1902 }
1903
1904 /* Get an address range which is currently unmapped.
1905  * For shmat() with addr=0.
1906  *
1907  * Ugly calling convention alert:
1908  * Return value with the low bits set means error value,
1909  * ie
1910  *      if (ret & ~PAGE_MASK)
1911  *              error = ret;
1912  *
1913  * This function "knows" that -ENOMEM has the bits set.
1914  */
1915 #ifndef HAVE_ARCH_UNMAPPED_AREA
1916 unsigned long
1917 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1918                 unsigned long len, unsigned long pgoff, unsigned long flags)
1919 {
1920         struct mm_struct *mm = current->mm;
1921         struct vm_area_struct *vma;
1922         struct vm_unmapped_area_info info;
1923
1924         if (len > TASK_SIZE - mmap_min_addr)
1925                 return -ENOMEM;
1926
1927         if (flags & MAP_FIXED)
1928                 return addr;
1929
1930         if (addr) {
1931                 addr = PAGE_ALIGN(addr);
1932                 vma = find_vma(mm, addr);
1933                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1934                     (!vma || addr + len <= vma->vm_start))
1935                         return addr;
1936         }
1937
1938         info.flags = 0;
1939         info.length = len;
1940         info.low_limit = mm->mmap_base;
1941         info.high_limit = TASK_SIZE;
1942         info.align_mask = 0;
1943         return vm_unmapped_area(&info);
1944 }
1945 #endif
1946
1947 /*
1948  * This mmap-allocator allocates new areas top-down from below the
1949  * stack's low limit (the base):
1950  */
1951 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1952 unsigned long
1953 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1954                           const unsigned long len, const unsigned long pgoff,
1955                           const unsigned long flags)
1956 {
1957         struct vm_area_struct *vma;
1958         struct mm_struct *mm = current->mm;
1959         unsigned long addr = addr0;
1960         struct vm_unmapped_area_info info;
1961
1962         /* requested length too big for entire address space */
1963         if (len > TASK_SIZE - mmap_min_addr)
1964                 return -ENOMEM;
1965
1966         if (flags & MAP_FIXED)
1967                 return addr;
1968
1969         /* requesting a specific address */
1970         if (addr) {
1971                 addr = PAGE_ALIGN(addr);
1972                 vma = find_vma(mm, addr);
1973                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1974                                 (!vma || addr + len <= vma->vm_start))
1975                         return addr;
1976         }
1977
1978         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1979         info.length = len;
1980         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1981         info.high_limit = mm->mmap_base;
1982         info.align_mask = 0;
1983         addr = vm_unmapped_area(&info);
1984
1985         /*
1986          * A failed mmap() very likely causes application failure,
1987          * so fall back to the bottom-up function here. This scenario
1988          * can happen with large stack limits and large mmap()
1989          * allocations.
1990          */
1991         if (addr & ~PAGE_MASK) {
1992                 VM_BUG_ON(addr != -ENOMEM);
1993                 info.flags = 0;
1994                 info.low_limit = TASK_UNMAPPED_BASE;
1995                 info.high_limit = TASK_SIZE;
1996                 addr = vm_unmapped_area(&info);
1997         }
1998
1999         return addr;
2000 }
2001 #endif
2002
2003 unsigned long
2004 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2005                 unsigned long pgoff, unsigned long flags)
2006 {
2007         unsigned long (*get_area)(struct file *, unsigned long,
2008                                   unsigned long, unsigned long, unsigned long);
2009
2010         unsigned long error = arch_mmap_check(addr, len, flags);
2011         if (error)
2012                 return error;
2013
2014         /* Careful about overflows.. */
2015         if (len > TASK_SIZE)
2016                 return -ENOMEM;
2017
2018         get_area = current->mm->get_unmapped_area;
2019         if (file && file->f_op->get_unmapped_area)
2020                 get_area = file->f_op->get_unmapped_area;
2021         addr = get_area(file, addr, len, pgoff, flags);
2022         if (IS_ERR_VALUE(addr))
2023                 return addr;
2024
2025         if (addr > TASK_SIZE - len)
2026                 return -ENOMEM;
2027         if (addr & ~PAGE_MASK)
2028                 return -EINVAL;
2029
2030         addr = arch_rebalance_pgtables(addr, len);
2031         error = security_mmap_addr(addr);
2032         return error ? error : addr;
2033 }
2034
2035 EXPORT_SYMBOL(get_unmapped_area);
2036
2037 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2038 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2039 {
2040         struct rb_node *rb_node;
2041         struct vm_area_struct *vma;
2042
2043         /* Check the cache first. */
2044         vma = vmacache_find(mm, addr);
2045         if (likely(vma))
2046                 return vma;
2047
2048         rb_node = mm->mm_rb.rb_node;
2049         vma = NULL;
2050
2051         while (rb_node) {
2052                 struct vm_area_struct *tmp;
2053
2054                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2055
2056                 if (tmp->vm_end > addr) {
2057                         vma = tmp;
2058                         if (tmp->vm_start <= addr)
2059                                 break;
2060                         rb_node = rb_node->rb_left;
2061                 } else
2062                         rb_node = rb_node->rb_right;
2063         }
2064
2065         if (vma)
2066                 vmacache_update(addr, vma);
2067         return vma;
2068 }
2069
2070 EXPORT_SYMBOL(find_vma);
2071
2072 /*
2073  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2074  */
2075 struct vm_area_struct *
2076 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2077                         struct vm_area_struct **pprev)
2078 {
2079         struct vm_area_struct *vma;
2080
2081         vma = find_vma(mm, addr);
2082         if (vma) {
2083                 *pprev = vma->vm_prev;
2084         } else {
2085                 struct rb_node *rb_node = mm->mm_rb.rb_node;
2086                 *pprev = NULL;
2087                 while (rb_node) {
2088                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2089                         rb_node = rb_node->rb_right;
2090                 }
2091         }
2092         return vma;
2093 }
2094
2095 /*
2096  * Verify that the stack growth is acceptable and
2097  * update accounting. This is shared with both the
2098  * grow-up and grow-down cases.
2099  */
2100 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2101 {
2102         struct mm_struct *mm = vma->vm_mm;
2103         struct rlimit *rlim = current->signal->rlim;
2104         unsigned long new_start, actual_size;
2105
2106         /* address space limit tests */
2107         if (!may_expand_vm(mm, grow))
2108                 return -ENOMEM;
2109
2110         /* Stack limit test */
2111         actual_size = size;
2112         if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN)))
2113                 actual_size -= PAGE_SIZE;
2114         if (actual_size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2115                 return -ENOMEM;
2116
2117         /* mlock limit tests */
2118         if (vma->vm_flags & VM_LOCKED) {
2119                 unsigned long locked;
2120                 unsigned long limit;
2121                 locked = mm->locked_vm + grow;
2122                 limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2123                 limit >>= PAGE_SHIFT;
2124                 if (locked > limit && !capable(CAP_IPC_LOCK))
2125                         return -ENOMEM;
2126         }
2127
2128         /* Check to ensure the stack will not grow into a hugetlb-only region */
2129         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2130                         vma->vm_end - size;
2131         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2132                 return -EFAULT;
2133
2134         /*
2135          * Overcommit..  This must be the final test, as it will
2136          * update security statistics.
2137          */
2138         if (security_vm_enough_memory_mm(mm, grow))
2139                 return -ENOMEM;
2140
2141         /* Ok, everything looks good - let it rip */
2142         if (vma->vm_flags & VM_LOCKED)
2143                 mm->locked_vm += grow;
2144         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
2145         return 0;
2146 }
2147
2148 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2149 /*
2150  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2151  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2152  */
2153 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2154 {
2155         int error;
2156
2157         if (!(vma->vm_flags & VM_GROWSUP))
2158                 return -EFAULT;
2159
2160         /*
2161          * We must make sure the anon_vma is allocated
2162          * so that the anon_vma locking is not a noop.
2163          */
2164         if (unlikely(anon_vma_prepare(vma)))
2165                 return -ENOMEM;
2166         vma_lock_anon_vma(vma);
2167
2168         /*
2169          * vma->vm_start/vm_end cannot change under us because the caller
2170          * is required to hold the mmap_sem in read mode.  We need the
2171          * anon_vma lock to serialize against concurrent expand_stacks.
2172          * Also guard against wrapping around to address 0.
2173          */
2174         if (address < PAGE_ALIGN(address+4))
2175                 address = PAGE_ALIGN(address+4);
2176         else {
2177                 vma_unlock_anon_vma(vma);
2178                 return -ENOMEM;
2179         }
2180         error = 0;
2181
2182         /* Somebody else might have raced and expanded it already */
2183         if (address > vma->vm_end) {
2184                 unsigned long size, grow;
2185
2186                 size = address - vma->vm_start;
2187                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2188
2189                 error = -ENOMEM;
2190                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2191                         error = acct_stack_growth(vma, size, grow);
2192                         if (!error) {
2193                                 /*
2194                                  * vma_gap_update() doesn't support concurrent
2195                                  * updates, but we only hold a shared mmap_sem
2196                                  * lock here, so we need to protect against
2197                                  * concurrent vma expansions.
2198                                  * vma_lock_anon_vma() doesn't help here, as
2199                                  * we don't guarantee that all growable vmas
2200                                  * in a mm share the same root anon vma.
2201                                  * So, we reuse mm->page_table_lock to guard
2202                                  * against concurrent vma expansions.
2203                                  */
2204                                 spin_lock(&vma->vm_mm->page_table_lock);
2205                                 anon_vma_interval_tree_pre_update_vma(vma);
2206                                 vma->vm_end = address;
2207                                 anon_vma_interval_tree_post_update_vma(vma);
2208                                 if (vma->vm_next)
2209                                         vma_gap_update(vma->vm_next);
2210                                 else
2211                                         vma->vm_mm->highest_vm_end = address;
2212                                 spin_unlock(&vma->vm_mm->page_table_lock);
2213
2214                                 perf_event_mmap(vma);
2215                         }
2216                 }
2217         }
2218         vma_unlock_anon_vma(vma);
2219         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2220         validate_mm(vma->vm_mm);
2221         return error;
2222 }
2223 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2224
2225 /*
2226  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2227  */
2228 int expand_downwards(struct vm_area_struct *vma,
2229                                    unsigned long address)
2230 {
2231         int error;
2232
2233         /*
2234          * We must make sure the anon_vma is allocated
2235          * so that the anon_vma locking is not a noop.
2236          */
2237         if (unlikely(anon_vma_prepare(vma)))
2238                 return -ENOMEM;
2239
2240         address &= PAGE_MASK;
2241         error = security_mmap_addr(address);
2242         if (error)
2243                 return error;
2244
2245         vma_lock_anon_vma(vma);
2246
2247         /*
2248          * vma->vm_start/vm_end cannot change under us because the caller
2249          * is required to hold the mmap_sem in read mode.  We need the
2250          * anon_vma lock to serialize against concurrent expand_stacks.
2251          */
2252
2253         /* Somebody else might have raced and expanded it already */
2254         if (address < vma->vm_start) {
2255                 unsigned long size, grow;
2256
2257                 size = vma->vm_end - address;
2258                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2259
2260                 error = -ENOMEM;
2261                 if (grow <= vma->vm_pgoff) {
2262                         error = acct_stack_growth(vma, size, grow);
2263                         if (!error) {
2264                                 /*
2265                                  * vma_gap_update() doesn't support concurrent
2266                                  * updates, but we only hold a shared mmap_sem
2267                                  * lock here, so we need to protect against
2268                                  * concurrent vma expansions.
2269                                  * vma_lock_anon_vma() doesn't help here, as
2270                                  * we don't guarantee that all growable vmas
2271                                  * in a mm share the same root anon vma.
2272                                  * So, we reuse mm->page_table_lock to guard
2273                                  * against concurrent vma expansions.
2274                                  */
2275                                 spin_lock(&vma->vm_mm->page_table_lock);
2276                                 anon_vma_interval_tree_pre_update_vma(vma);
2277                                 vma->vm_start = address;
2278                                 vma->vm_pgoff -= grow;
2279                                 anon_vma_interval_tree_post_update_vma(vma);
2280                                 vma_gap_update(vma);
2281                                 spin_unlock(&vma->vm_mm->page_table_lock);
2282
2283                                 perf_event_mmap(vma);
2284                         }
2285                 }
2286         }
2287         vma_unlock_anon_vma(vma);
2288         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2289         validate_mm(vma->vm_mm);
2290         return error;
2291 }
2292
2293 /*
2294  * Note how expand_stack() refuses to expand the stack all the way to
2295  * abut the next virtual mapping, *unless* that mapping itself is also
2296  * a stack mapping. We want to leave room for a guard page, after all
2297  * (the guard page itself is not added here, that is done by the
2298  * actual page faulting logic)
2299  *
2300  * This matches the behavior of the guard page logic (see mm/memory.c:
2301  * check_stack_guard_page()), which only allows the guard page to be
2302  * removed under these circumstances.
2303  */
2304 #ifdef CONFIG_STACK_GROWSUP
2305 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2306 {
2307         struct vm_area_struct *next;
2308
2309         address &= PAGE_MASK;
2310         next = vma->vm_next;
2311         if (next && next->vm_start == address + PAGE_SIZE) {
2312                 if (!(next->vm_flags & VM_GROWSUP))
2313                         return -ENOMEM;
2314         }
2315         return expand_upwards(vma, address);
2316 }
2317
2318 struct vm_area_struct *
2319 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2320 {
2321         struct vm_area_struct *vma, *prev;
2322
2323         addr &= PAGE_MASK;
2324         vma = find_vma_prev(mm, addr, &prev);
2325         if (vma && (vma->vm_start <= addr))
2326                 return vma;
2327         if (!prev || expand_stack(prev, addr))
2328                 return NULL;
2329         if (prev->vm_flags & VM_LOCKED)
2330                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2331         return prev;
2332 }
2333 #else
2334 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2335 {
2336         struct vm_area_struct *prev;
2337
2338         address &= PAGE_MASK;
2339         prev = vma->vm_prev;
2340         if (prev && prev->vm_end == address) {
2341                 if (!(prev->vm_flags & VM_GROWSDOWN))
2342                         return -ENOMEM;
2343         }
2344         return expand_downwards(vma, address);
2345 }
2346
2347 struct vm_area_struct *
2348 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2349 {
2350         struct vm_area_struct *vma;
2351         unsigned long start;
2352
2353         addr &= PAGE_MASK;
2354         vma = find_vma(mm, addr);
2355         if (!vma)
2356                 return NULL;
2357         if (vma->vm_start <= addr)
2358                 return vma;
2359         if (!(vma->vm_flags & VM_GROWSDOWN))
2360                 return NULL;
2361         start = vma->vm_start;
2362         if (expand_stack(vma, addr))
2363                 return NULL;
2364         if (vma->vm_flags & VM_LOCKED)
2365                 populate_vma_page_range(vma, addr, start, NULL);
2366         return vma;
2367 }
2368 #endif
2369
2370 EXPORT_SYMBOL_GPL(find_extend_vma);
2371
2372 /*
2373  * Ok - we have the memory areas we should free on the vma list,
2374  * so release them, and do the vma updates.
2375  *
2376  * Called with the mm semaphore held.
2377  */
2378 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2379 {
2380         unsigned long nr_accounted = 0;
2381
2382         /* Update high watermark before we lower total_vm */
2383         update_hiwater_vm(mm);
2384         do {
2385                 long nrpages = vma_pages(vma);
2386
2387                 if (vma->vm_flags & VM_ACCOUNT)
2388                         nr_accounted += nrpages;
2389                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2390                 vma = remove_vma(vma);
2391         } while (vma);
2392         vm_unacct_memory(nr_accounted);
2393         validate_mm(mm);
2394 }
2395
2396 /*
2397  * Get rid of page table information in the indicated region.
2398  *
2399  * Called with the mm semaphore held.
2400  */
2401 static void unmap_region(struct mm_struct *mm,
2402                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2403                 unsigned long start, unsigned long end)
2404 {
2405         struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2406         struct mmu_gather tlb;
2407
2408         lru_add_drain();
2409         tlb_gather_mmu(&tlb, mm, start, end);
2410         update_hiwater_rss(mm);
2411         unmap_vmas(&tlb, vma, start, end);
2412         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2413                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2414         tlb_finish_mmu(&tlb, start, end);
2415 }
2416
2417 /*
2418  * Create a list of vma's touched by the unmap, removing them from the mm's
2419  * vma list as we go..
2420  */
2421 static void
2422 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2423         struct vm_area_struct *prev, unsigned long end)
2424 {
2425         struct vm_area_struct **insertion_point;
2426         struct vm_area_struct *tail_vma = NULL;
2427
2428         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2429         vma->vm_prev = NULL;
2430         do {
2431                 vma_rb_erase(vma, &mm->mm_rb);
2432                 mm->map_count--;
2433                 tail_vma = vma;
2434                 vma = vma->vm_next;
2435         } while (vma && vma->vm_start < end);
2436         *insertion_point = vma;
2437         if (vma) {
2438                 vma->vm_prev = prev;
2439                 vma_gap_update(vma);
2440         } else
2441                 mm->highest_vm_end = prev ? prev->vm_end : 0;
2442         tail_vma->vm_next = NULL;
2443
2444         /* Kill the cache */
2445         vmacache_invalidate(mm);
2446 }
2447
2448 /*
2449  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2450  * munmap path where it doesn't make sense to fail.
2451  */
2452 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2453               unsigned long addr, int new_below)
2454 {
2455         struct vm_area_struct *new;
2456         int err;
2457
2458         if (is_vm_hugetlb_page(vma) && (addr &
2459                                         ~(huge_page_mask(hstate_vma(vma)))))
2460                 return -EINVAL;
2461
2462         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2463         if (!new)
2464                 return -ENOMEM;
2465
2466         /* most fields are the same, copy all, and then fixup */
2467         *new = *vma;
2468
2469         INIT_LIST_HEAD(&new->anon_vma_chain);
2470
2471         if (new_below)
2472                 new->vm_end = addr;
2473         else {
2474                 new->vm_start = addr;
2475                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2476         }
2477
2478         err = vma_dup_policy(vma, new);
2479         if (err)
2480                 goto out_free_vma;
2481
2482         err = anon_vma_clone(new, vma);
2483         if (err)
2484                 goto out_free_mpol;
2485
2486         if (new->vm_file)
2487                 get_file(new->vm_file);
2488
2489         if (new->vm_ops && new->vm_ops->open)
2490                 new->vm_ops->open(new);
2491
2492         if (new_below)
2493                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2494                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2495         else
2496                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2497
2498         /* Success. */
2499         if (!err)
2500                 return 0;
2501
2502         /* Clean everything up if vma_adjust failed. */
2503         if (new->vm_ops && new->vm_ops->close)
2504                 new->vm_ops->close(new);
2505         if (new->vm_file)
2506                 fput(new->vm_file);
2507         unlink_anon_vmas(new);
2508  out_free_mpol:
2509         mpol_put(vma_policy(new));
2510  out_free_vma:
2511         kmem_cache_free(vm_area_cachep, new);
2512         return err;
2513 }
2514
2515 /*
2516  * Split a vma into two pieces at address 'addr', a new vma is allocated
2517  * either for the first part or the tail.
2518  */
2519 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2520               unsigned long addr, int new_below)
2521 {
2522         if (mm->map_count >= sysctl_max_map_count)
2523                 return -ENOMEM;
2524
2525         return __split_vma(mm, vma, addr, new_below);
2526 }
2527
2528 /* Munmap is split into 2 main parts -- this part which finds
2529  * what needs doing, and the areas themselves, which do the
2530  * work.  This now handles partial unmappings.
2531  * Jeremy Fitzhardinge <jeremy@goop.org>
2532  */
2533 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2534 {
2535         unsigned long end;
2536         struct vm_area_struct *vma, *prev, *last;
2537
2538         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2539                 return -EINVAL;
2540
2541         len = PAGE_ALIGN(len);
2542         if (len == 0)
2543                 return -EINVAL;
2544
2545         /* Find the first overlapping VMA */
2546         vma = find_vma(mm, start);
2547         if (!vma)
2548                 return 0;
2549         prev = vma->vm_prev;
2550         /* we have  start < vma->vm_end  */
2551
2552         /* if it doesn't overlap, we have nothing.. */
2553         end = start + len;
2554         if (vma->vm_start >= end)
2555                 return 0;
2556
2557         /*
2558          * If we need to split any vma, do it now to save pain later.
2559          *
2560          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2561          * unmapped vm_area_struct will remain in use: so lower split_vma
2562          * places tmp vma above, and higher split_vma places tmp vma below.
2563          */
2564         if (start > vma->vm_start) {
2565                 int error;
2566
2567                 /*
2568                  * Make sure that map_count on return from munmap() will
2569                  * not exceed its limit; but let map_count go just above
2570                  * its limit temporarily, to help free resources as expected.
2571                  */
2572                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2573                         return -ENOMEM;
2574
2575                 error = __split_vma(mm, vma, start, 0);
2576                 if (error)
2577                         return error;
2578                 prev = vma;
2579         }
2580
2581         /* Does it split the last one? */
2582         last = find_vma(mm, end);
2583         if (last && end > last->vm_start) {
2584                 int error = __split_vma(mm, last, end, 1);
2585                 if (error)
2586                         return error;
2587         }
2588         vma = prev ? prev->vm_next : mm->mmap;
2589
2590         /*
2591          * unlock any mlock()ed ranges before detaching vmas
2592          */
2593         if (mm->locked_vm) {
2594                 struct vm_area_struct *tmp = vma;
2595                 while (tmp && tmp->vm_start < end) {
2596                         if (tmp->vm_flags & VM_LOCKED) {
2597                                 mm->locked_vm -= vma_pages(tmp);
2598                                 munlock_vma_pages_all(tmp);
2599                         }
2600                         tmp = tmp->vm_next;
2601                 }
2602         }
2603
2604         /*
2605          * Remove the vma's, and unmap the actual pages
2606          */
2607         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2608         unmap_region(mm, vma, prev, start, end);
2609
2610         arch_unmap(mm, vma, start, end);
2611
2612         /* Fix up all other VM information */
2613         remove_vma_list(mm, vma);
2614
2615         return 0;
2616 }
2617
2618 int vm_munmap(unsigned long start, size_t len)
2619 {
2620         int ret;
2621         struct mm_struct *mm = current->mm;
2622
2623         down_write(&mm->mmap_sem);
2624         ret = do_munmap(mm, start, len);
2625         up_write(&mm->mmap_sem);
2626         return ret;
2627 }
2628 EXPORT_SYMBOL(vm_munmap);
2629
2630 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2631 {
2632         profile_munmap(addr);
2633         return vm_munmap(addr, len);
2634 }
2635
2636
2637 /*
2638  * Emulation of deprecated remap_file_pages() syscall.
2639  */
2640 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2641                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2642 {
2643
2644         struct mm_struct *mm = current->mm;
2645         struct vm_area_struct *vma;
2646         unsigned long populate = 0;
2647         unsigned long ret = -EINVAL;
2648         struct file *file;
2649
2650         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. "
2651                         "See Documentation/vm/remap_file_pages.txt.\n",
2652                         current->comm, current->pid);
2653
2654         if (prot)
2655                 return ret;
2656         start = start & PAGE_MASK;
2657         size = size & PAGE_MASK;
2658
2659         if (start + size <= start)
2660                 return ret;
2661
2662         /* Does pgoff wrap? */
2663         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2664                 return ret;
2665
2666         down_write(&mm->mmap_sem);
2667         vma = find_vma(mm, start);
2668
2669         if (!vma || !(vma->vm_flags & VM_SHARED))
2670                 goto out;
2671
2672         if (start < vma->vm_start || start + size > vma->vm_end)
2673                 goto out;
2674
2675         if (pgoff == linear_page_index(vma, start)) {
2676                 ret = 0;
2677                 goto out;
2678         }
2679
2680         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2681         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2682         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2683
2684         flags &= MAP_NONBLOCK;
2685         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2686         if (vma->vm_flags & VM_LOCKED) {
2687                 flags |= MAP_LOCKED;
2688                 /* drop PG_Mlocked flag for over-mapped range */
2689                 munlock_vma_pages_range(vma, start, start + size);
2690         }
2691
2692         file = get_file(vma->vm_file);
2693         ret = do_mmap_pgoff(vma->vm_file, start, size,
2694                         prot, flags, pgoff, &populate);
2695         fput(file);
2696 out:
2697         up_write(&mm->mmap_sem);
2698         if (populate)
2699                 mm_populate(ret, populate);
2700         if (!IS_ERR_VALUE(ret))
2701                 ret = 0;
2702         return ret;
2703 }
2704
2705 static inline void verify_mm_writelocked(struct mm_struct *mm)
2706 {
2707 #ifdef CONFIG_DEBUG_VM
2708         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2709                 WARN_ON(1);
2710                 up_read(&mm->mmap_sem);
2711         }
2712 #endif
2713 }
2714
2715 /*
2716  *  this is really a simplified "do_mmap".  it only handles
2717  *  anonymous maps.  eventually we may be able to do some
2718  *  brk-specific accounting here.
2719  */
2720 static unsigned long do_brk(unsigned long addr, unsigned long len)
2721 {
2722         struct mm_struct *mm = current->mm;
2723         struct vm_area_struct *vma, *prev;
2724         unsigned long flags;
2725         struct rb_node **rb_link, *rb_parent;
2726         pgoff_t pgoff = addr >> PAGE_SHIFT;
2727         int error;
2728
2729         len = PAGE_ALIGN(len);
2730         if (!len)
2731                 return addr;
2732
2733         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2734
2735         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2736         if (error & ~PAGE_MASK)
2737                 return error;
2738
2739         error = mlock_future_check(mm, mm->def_flags, len);
2740         if (error)
2741                 return error;
2742
2743         /*
2744          * mm->mmap_sem is required to protect against another thread
2745          * changing the mappings in case we sleep.
2746          */
2747         verify_mm_writelocked(mm);
2748
2749         /*
2750          * Clear old maps.  this also does some error checking for us
2751          */
2752         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2753                               &rb_parent)) {
2754                 if (do_munmap(mm, addr, len))
2755                         return -ENOMEM;
2756         }
2757
2758         /* Check against address space limits *after* clearing old maps... */
2759         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2760                 return -ENOMEM;
2761
2762         if (mm->map_count > sysctl_max_map_count)
2763                 return -ENOMEM;
2764
2765         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2766                 return -ENOMEM;
2767
2768         /* Can we just expand an old private anonymous mapping? */
2769         vma = vma_merge(mm, prev, addr, addr + len, flags,
2770                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2771         if (vma)
2772                 goto out;
2773
2774         /*
2775          * create a vma struct for an anonymous mapping
2776          */
2777         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2778         if (!vma) {
2779                 vm_unacct_memory(len >> PAGE_SHIFT);
2780                 return -ENOMEM;
2781         }
2782
2783         INIT_LIST_HEAD(&vma->anon_vma_chain);
2784         vma->vm_mm = mm;
2785         vma->vm_start = addr;
2786         vma->vm_end = addr + len;
2787         vma->vm_pgoff = pgoff;
2788         vma->vm_flags = flags;
2789         vma->vm_page_prot = vm_get_page_prot(flags);
2790         vma_link(mm, vma, prev, rb_link, rb_parent);
2791 out:
2792         perf_event_mmap(vma);
2793         mm->total_vm += len >> PAGE_SHIFT;
2794         if (flags & VM_LOCKED)
2795                 mm->locked_vm += (len >> PAGE_SHIFT);
2796         vma->vm_flags |= VM_SOFTDIRTY;
2797         return addr;
2798 }
2799
2800 unsigned long vm_brk(unsigned long addr, unsigned long len)
2801 {
2802         struct mm_struct *mm = current->mm;
2803         unsigned long ret;
2804         bool populate;
2805
2806         down_write(&mm->mmap_sem);
2807         ret = do_brk(addr, len);
2808         populate = ((mm->def_flags & VM_LOCKED) != 0);
2809         up_write(&mm->mmap_sem);
2810         if (populate)
2811                 mm_populate(addr, len);
2812         return ret;
2813 }
2814 EXPORT_SYMBOL(vm_brk);
2815
2816 /* Release all mmaps. */
2817 void exit_mmap(struct mm_struct *mm)
2818 {
2819         struct mmu_gather tlb;
2820         struct vm_area_struct *vma;
2821         unsigned long nr_accounted = 0;
2822
2823         /* mm's last user has gone, and its about to be pulled down */
2824         mmu_notifier_release(mm);
2825
2826         if (mm->locked_vm) {
2827                 vma = mm->mmap;
2828                 while (vma) {
2829                         if (vma->vm_flags & VM_LOCKED)
2830                                 munlock_vma_pages_all(vma);
2831                         vma = vma->vm_next;
2832                 }
2833         }
2834
2835         arch_exit_mmap(mm);
2836
2837         vma = mm->mmap;
2838         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2839                 return;
2840
2841         lru_add_drain();
2842         flush_cache_mm(mm);
2843         tlb_gather_mmu(&tlb, mm, 0, -1);
2844         /* update_hiwater_rss(mm) here? but nobody should be looking */
2845         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2846         unmap_vmas(&tlb, vma, 0, -1);
2847
2848         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2849         tlb_finish_mmu(&tlb, 0, -1);
2850
2851         /*
2852          * Walk the list again, actually closing and freeing it,
2853          * with preemption enabled, without holding any MM locks.
2854          */
2855         while (vma) {
2856                 if (vma->vm_flags & VM_ACCOUNT)
2857                         nr_accounted += vma_pages(vma);
2858                 vma = remove_vma(vma);
2859         }
2860         vm_unacct_memory(nr_accounted);
2861 }
2862
2863 /* Insert vm structure into process list sorted by address
2864  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2865  * then i_mmap_rwsem is taken here.
2866  */
2867 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2868 {
2869         struct vm_area_struct *prev;
2870         struct rb_node **rb_link, *rb_parent;
2871
2872         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2873                            &prev, &rb_link, &rb_parent))
2874                 return -ENOMEM;
2875         if ((vma->vm_flags & VM_ACCOUNT) &&
2876              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2877                 return -ENOMEM;
2878
2879         /*
2880          * The vm_pgoff of a purely anonymous vma should be irrelevant
2881          * until its first write fault, when page's anon_vma and index
2882          * are set.  But now set the vm_pgoff it will almost certainly
2883          * end up with (unless mremap moves it elsewhere before that
2884          * first wfault), so /proc/pid/maps tells a consistent story.
2885          *
2886          * By setting it to reflect the virtual start address of the
2887          * vma, merges and splits can happen in a seamless way, just
2888          * using the existing file pgoff checks and manipulations.
2889          * Similarly in do_mmap_pgoff and in do_brk.
2890          */
2891         if (vma_is_anonymous(vma)) {
2892                 BUG_ON(vma->anon_vma);
2893                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2894         }
2895
2896         vma_link(mm, vma, prev, rb_link, rb_parent);
2897         return 0;
2898 }
2899
2900 /*
2901  * Copy the vma structure to a new location in the same mm,
2902  * prior to moving page table entries, to effect an mremap move.
2903  */
2904 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2905         unsigned long addr, unsigned long len, pgoff_t pgoff,
2906         bool *need_rmap_locks)
2907 {
2908         struct vm_area_struct *vma = *vmap;
2909         unsigned long vma_start = vma->vm_start;
2910         struct mm_struct *mm = vma->vm_mm;
2911         struct vm_area_struct *new_vma, *prev;
2912         struct rb_node **rb_link, *rb_parent;
2913         bool faulted_in_anon_vma = true;
2914
2915         /*
2916          * If anonymous vma has not yet been faulted, update new pgoff
2917          * to match new location, to increase its chance of merging.
2918          */
2919         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
2920                 pgoff = addr >> PAGE_SHIFT;
2921                 faulted_in_anon_vma = false;
2922         }
2923
2924         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2925                 return NULL;    /* should never get here */
2926         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2927                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
2928                             vma->vm_userfaultfd_ctx);
2929         if (new_vma) {
2930                 /*
2931                  * Source vma may have been merged into new_vma
2932                  */
2933                 if (unlikely(vma_start >= new_vma->vm_start &&
2934                              vma_start < new_vma->vm_end)) {
2935                         /*
2936                          * The only way we can get a vma_merge with
2937                          * self during an mremap is if the vma hasn't
2938                          * been faulted in yet and we were allowed to
2939                          * reset the dst vma->vm_pgoff to the
2940                          * destination address of the mremap to allow
2941                          * the merge to happen. mremap must change the
2942                          * vm_pgoff linearity between src and dst vmas
2943                          * (in turn preventing a vma_merge) to be
2944                          * safe. It is only safe to keep the vm_pgoff
2945                          * linear if there are no pages mapped yet.
2946                          */
2947                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
2948                         *vmap = vma = new_vma;
2949                 }
2950                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2951         } else {
2952                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2953                 if (!new_vma)
2954                         goto out;
2955                 *new_vma = *vma;
2956                 new_vma->vm_start = addr;
2957                 new_vma->vm_end = addr + len;
2958                 new_vma->vm_pgoff = pgoff;
2959                 if (vma_dup_policy(vma, new_vma))
2960                         goto out_free_vma;
2961                 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2962                 if (anon_vma_clone(new_vma, vma))
2963                         goto out_free_mempol;
2964                 if (new_vma->vm_file)
2965                         get_file(new_vma->vm_file);
2966                 if (new_vma->vm_ops && new_vma->vm_ops->open)
2967                         new_vma->vm_ops->open(new_vma);
2968                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2969                 *need_rmap_locks = false;
2970         }
2971         return new_vma;
2972
2973 out_free_mempol:
2974         mpol_put(vma_policy(new_vma));
2975 out_free_vma:
2976         kmem_cache_free(vm_area_cachep, new_vma);
2977 out:
2978         return NULL;
2979 }
2980
2981 /*
2982  * Return true if the calling process may expand its vm space by the passed
2983  * number of pages
2984  */
2985 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2986 {
2987         unsigned long cur = mm->total_vm;       /* pages */
2988         unsigned long lim;
2989
2990         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2991
2992         if (cur + npages > lim)
2993                 return 0;
2994         return 1;
2995 }
2996
2997 static int special_mapping_fault(struct vm_area_struct *vma,
2998                                  struct vm_fault *vmf);
2999
3000 /*
3001  * Having a close hook prevents vma merging regardless of flags.
3002  */
3003 static void special_mapping_close(struct vm_area_struct *vma)
3004 {
3005 }
3006
3007 static const char *special_mapping_name(struct vm_area_struct *vma)
3008 {
3009         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3010 }
3011
3012 static const struct vm_operations_struct special_mapping_vmops = {
3013         .close = special_mapping_close,
3014         .fault = special_mapping_fault,
3015         .name = special_mapping_name,
3016 };
3017
3018 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3019         .close = special_mapping_close,
3020         .fault = special_mapping_fault,
3021 };
3022
3023 static int special_mapping_fault(struct vm_area_struct *vma,
3024                                 struct vm_fault *vmf)
3025 {
3026         pgoff_t pgoff;
3027         struct page **pages;
3028
3029         if (vma->vm_ops == &legacy_special_mapping_vmops)
3030                 pages = vma->vm_private_data;
3031         else
3032                 pages = ((struct vm_special_mapping *)vma->vm_private_data)->
3033                         pages;
3034
3035         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3036                 pgoff--;
3037
3038         if (*pages) {
3039                 struct page *page = *pages;
3040                 get_page(page);
3041                 vmf->page = page;
3042                 return 0;
3043         }
3044
3045         return VM_FAULT_SIGBUS;
3046 }
3047
3048 static struct vm_area_struct *__install_special_mapping(
3049         struct mm_struct *mm,
3050         unsigned long addr, unsigned long len,
3051         unsigned long vm_flags, const struct vm_operations_struct *ops,
3052         void *priv)
3053 {
3054         int ret;
3055         struct vm_area_struct *vma;
3056
3057         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3058         if (unlikely(vma == NULL))
3059                 return ERR_PTR(-ENOMEM);
3060
3061         INIT_LIST_HEAD(&vma->anon_vma_chain);
3062         vma->vm_mm = mm;
3063         vma->vm_start = addr;
3064         vma->vm_end = addr + len;
3065
3066         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3067         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3068
3069         vma->vm_ops = ops;
3070         vma->vm_private_data = priv;
3071
3072         ret = insert_vm_struct(mm, vma);
3073         if (ret)
3074                 goto out;
3075
3076         mm->total_vm += len >> PAGE_SHIFT;
3077
3078         perf_event_mmap(vma);
3079
3080         return vma;
3081
3082 out:
3083         kmem_cache_free(vm_area_cachep, vma);
3084         return ERR_PTR(ret);
3085 }
3086
3087 /*
3088  * Called with mm->mmap_sem held for writing.
3089  * Insert a new vma covering the given region, with the given flags.
3090  * Its pages are supplied by the given array of struct page *.
3091  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3092  * The region past the last page supplied will always produce SIGBUS.
3093  * The array pointer and the pages it points to are assumed to stay alive
3094  * for as long as this mapping might exist.
3095  */
3096 struct vm_area_struct *_install_special_mapping(
3097         struct mm_struct *mm,
3098         unsigned long addr, unsigned long len,
3099         unsigned long vm_flags, const struct vm_special_mapping *spec)
3100 {
3101         return __install_special_mapping(mm, addr, len, vm_flags,
3102                                          &special_mapping_vmops, (void *)spec);
3103 }
3104
3105 int install_special_mapping(struct mm_struct *mm,
3106                             unsigned long addr, unsigned long len,
3107                             unsigned long vm_flags, struct page **pages)
3108 {
3109         struct vm_area_struct *vma = __install_special_mapping(
3110                 mm, addr, len, vm_flags, &legacy_special_mapping_vmops,
3111                 (void *)pages);
3112
3113         return PTR_ERR_OR_ZERO(vma);
3114 }
3115
3116 static DEFINE_MUTEX(mm_all_locks_mutex);
3117
3118 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3119 {
3120         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3121                 /*
3122                  * The LSB of head.next can't change from under us
3123                  * because we hold the mm_all_locks_mutex.
3124                  */
3125                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3126                 /*
3127                  * We can safely modify head.next after taking the
3128                  * anon_vma->root->rwsem. If some other vma in this mm shares
3129                  * the same anon_vma we won't take it again.
3130                  *
3131                  * No need of atomic instructions here, head.next
3132                  * can't change from under us thanks to the
3133                  * anon_vma->root->rwsem.
3134                  */
3135                 if (__test_and_set_bit(0, (unsigned long *)
3136                                        &anon_vma->root->rb_root.rb_node))
3137                         BUG();
3138         }
3139 }
3140
3141 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3142 {
3143         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3144                 /*
3145                  * AS_MM_ALL_LOCKS can't change from under us because
3146                  * we hold the mm_all_locks_mutex.
3147                  *
3148                  * Operations on ->flags have to be atomic because
3149                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3150                  * mm_all_locks_mutex, there may be other cpus
3151                  * changing other bitflags in parallel to us.
3152                  */
3153                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3154                         BUG();
3155                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3156         }
3157 }
3158
3159 /*
3160  * This operation locks against the VM for all pte/vma/mm related
3161  * operations that could ever happen on a certain mm. This includes
3162  * vmtruncate, try_to_unmap, and all page faults.
3163  *
3164  * The caller must take the mmap_sem in write mode before calling
3165  * mm_take_all_locks(). The caller isn't allowed to release the
3166  * mmap_sem until mm_drop_all_locks() returns.
3167  *
3168  * mmap_sem in write mode is required in order to block all operations
3169  * that could modify pagetables and free pages without need of
3170  * altering the vma layout. It's also needed in write mode to avoid new
3171  * anon_vmas to be associated with existing vmas.
3172  *
3173  * A single task can't take more than one mm_take_all_locks() in a row
3174  * or it would deadlock.
3175  *
3176  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3177  * mapping->flags avoid to take the same lock twice, if more than one
3178  * vma in this mm is backed by the same anon_vma or address_space.
3179  *
3180  * We can take all the locks in random order because the VM code
3181  * taking i_mmap_rwsem or anon_vma->rwsem outside the mmap_sem never
3182  * takes more than one of them in a row. Secondly we're protected
3183  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3184  *
3185  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3186  * that may have to take thousand of locks.
3187  *
3188  * mm_take_all_locks() can fail if it's interrupted by signals.
3189  */
3190 int mm_take_all_locks(struct mm_struct *mm)
3191 {
3192         struct vm_area_struct *vma;
3193         struct anon_vma_chain *avc;
3194
3195         BUG_ON(down_read_trylock(&mm->mmap_sem));
3196
3197         mutex_lock(&mm_all_locks_mutex);
3198
3199         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3200                 if (signal_pending(current))
3201                         goto out_unlock;
3202                 if (vma->vm_file && vma->vm_file->f_mapping)
3203                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3204         }
3205
3206         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3207                 if (signal_pending(current))
3208                         goto out_unlock;
3209                 if (vma->anon_vma)
3210                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3211                                 vm_lock_anon_vma(mm, avc->anon_vma);
3212         }
3213
3214         return 0;
3215
3216 out_unlock:
3217         mm_drop_all_locks(mm);
3218         return -EINTR;
3219 }
3220
3221 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3222 {
3223         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3224                 /*
3225                  * The LSB of head.next can't change to 0 from under
3226                  * us because we hold the mm_all_locks_mutex.
3227                  *
3228                  * We must however clear the bitflag before unlocking
3229                  * the vma so the users using the anon_vma->rb_root will
3230                  * never see our bitflag.
3231                  *
3232                  * No need of atomic instructions here, head.next
3233                  * can't change from under us until we release the
3234                  * anon_vma->root->rwsem.
3235                  */
3236                 if (!__test_and_clear_bit(0, (unsigned long *)
3237                                           &anon_vma->root->rb_root.rb_node))
3238                         BUG();
3239                 anon_vma_unlock_write(anon_vma);
3240         }
3241 }
3242
3243 static void vm_unlock_mapping(struct address_space *mapping)
3244 {
3245         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3246                 /*
3247                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3248                  * because we hold the mm_all_locks_mutex.
3249                  */
3250                 i_mmap_unlock_write(mapping);
3251                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3252                                         &mapping->flags))
3253                         BUG();
3254         }
3255 }
3256
3257 /*
3258  * The mmap_sem cannot be released by the caller until
3259  * mm_drop_all_locks() returns.
3260  */
3261 void mm_drop_all_locks(struct mm_struct *mm)
3262 {
3263         struct vm_area_struct *vma;
3264         struct anon_vma_chain *avc;
3265
3266         BUG_ON(down_read_trylock(&mm->mmap_sem));
3267         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3268
3269         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3270                 if (vma->anon_vma)
3271                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3272                                 vm_unlock_anon_vma(avc->anon_vma);
3273                 if (vma->vm_file && vma->vm_file->f_mapping)
3274                         vm_unlock_mapping(vma->vm_file->f_mapping);
3275         }
3276
3277         mutex_unlock(&mm_all_locks_mutex);
3278 }
3279
3280 /*
3281  * initialise the VMA slab
3282  */
3283 void __init mmap_init(void)
3284 {
3285         int ret;
3286
3287         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3288         VM_BUG_ON(ret);
3289 }
3290
3291 /*
3292  * Initialise sysctl_user_reserve_kbytes.
3293  *
3294  * This is intended to prevent a user from starting a single memory hogging
3295  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3296  * mode.
3297  *
3298  * The default value is min(3% of free memory, 128MB)
3299  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3300  */
3301 static int init_user_reserve(void)
3302 {
3303         unsigned long free_kbytes;
3304
3305         free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3306
3307         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3308         return 0;
3309 }
3310 subsys_initcall(init_user_reserve);
3311
3312 /*
3313  * Initialise sysctl_admin_reserve_kbytes.
3314  *
3315  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3316  * to log in and kill a memory hogging process.
3317  *
3318  * Systems with more than 256MB will reserve 8MB, enough to recover
3319  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3320  * only reserve 3% of free pages by default.
3321  */
3322 static int init_admin_reserve(void)
3323 {
3324         unsigned long free_kbytes;
3325
3326         free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3327
3328         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3329         return 0;
3330 }
3331 subsys_initcall(init_admin_reserve);
3332
3333 /*
3334  * Reinititalise user and admin reserves if memory is added or removed.
3335  *
3336  * The default user reserve max is 128MB, and the default max for the
3337  * admin reserve is 8MB. These are usually, but not always, enough to
3338  * enable recovery from a memory hogging process using login/sshd, a shell,
3339  * and tools like top. It may make sense to increase or even disable the
3340  * reserve depending on the existence of swap or variations in the recovery
3341  * tools. So, the admin may have changed them.
3342  *
3343  * If memory is added and the reserves have been eliminated or increased above
3344  * the default max, then we'll trust the admin.
3345  *
3346  * If memory is removed and there isn't enough free memory, then we
3347  * need to reset the reserves.
3348  *
3349  * Otherwise keep the reserve set by the admin.
3350  */
3351 static int reserve_mem_notifier(struct notifier_block *nb,
3352                              unsigned long action, void *data)
3353 {
3354         unsigned long tmp, free_kbytes;
3355
3356         switch (action) {
3357         case MEM_ONLINE:
3358                 /* Default max is 128MB. Leave alone if modified by operator. */
3359                 tmp = sysctl_user_reserve_kbytes;
3360                 if (0 < tmp && tmp < (1UL << 17))
3361                         init_user_reserve();
3362
3363                 /* Default max is 8MB.  Leave alone if modified by operator. */
3364                 tmp = sysctl_admin_reserve_kbytes;
3365                 if (0 < tmp && tmp < (1UL << 13))
3366                         init_admin_reserve();
3367
3368                 break;
3369         case MEM_OFFLINE:
3370                 free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3371
3372                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3373                         init_user_reserve();
3374                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3375                                 sysctl_user_reserve_kbytes);
3376                 }
3377
3378                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3379                         init_admin_reserve();
3380                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3381                                 sysctl_admin_reserve_kbytes);
3382                 }
3383                 break;
3384         default:
3385                 break;
3386         }
3387         return NOTIFY_OK;
3388 }
3389
3390 static struct notifier_block reserve_mem_nb = {
3391         .notifier_call = reserve_mem_notifier,
3392 };
3393
3394 static int __meminit init_reserve_notifier(void)
3395 {
3396         if (register_hotmemory_notifier(&reserve_mem_nb))
3397                 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3398
3399         return 0;
3400 }
3401 subsys_initcall(init_reserve_notifier);