mm: always print RLIMIT_DATA warning
[muen/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/shmem_fs.h>
29 #include <linux/profile.h>
30 #include <linux/export.h>
31 #include <linux/mount.h>
32 #include <linux/mempolicy.h>
33 #include <linux/rmap.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/mmdebug.h>
36 #include <linux/perf_event.h>
37 #include <linux/audit.h>
38 #include <linux/khugepaged.h>
39 #include <linux/uprobes.h>
40 #include <linux/rbtree_augmented.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/moduleparam.h>
46 #include <linux/pkeys.h>
47 #include <linux/oom.h>
48
49 #include <linux/uaccess.h>
50 #include <asm/cacheflush.h>
51 #include <asm/tlb.h>
52 #include <asm/mmu_context.h>
53
54 #include "internal.h"
55
56 #ifndef arch_mmap_check
57 #define arch_mmap_check(addr, len, flags)       (0)
58 #endif
59
60 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
61 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
62 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
63 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
64 #endif
65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
66 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
67 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
68 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
69 #endif
70
71 static bool ignore_rlimit_data;
72 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
73
74 static void unmap_region(struct mm_struct *mm,
75                 struct vm_area_struct *vma, struct vm_area_struct *prev,
76                 unsigned long start, unsigned long end);
77
78 /* description of effects of mapping type and prot in current implementation.
79  * this is due to the limited x86 page protection hardware.  The expected
80  * behavior is in parens:
81  *
82  * map_type     prot
83  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
84  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
85  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
86  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
87  *
88  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
89  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
90  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
91  *
92  * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
93  * MAP_PRIVATE:
94  *                                                              r: (no) no
95  *                                                              w: (no) no
96  *                                                              x: (yes) yes
97  */
98 pgprot_t protection_map[16] __ro_after_init = {
99         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
100         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
101 };
102
103 pgprot_t vm_get_page_prot(unsigned long vm_flags)
104 {
105         return __pgprot(pgprot_val(protection_map[vm_flags &
106                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
107                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
108 }
109 EXPORT_SYMBOL(vm_get_page_prot);
110
111 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
112 {
113         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
114 }
115
116 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
117 void vma_set_page_prot(struct vm_area_struct *vma)
118 {
119         unsigned long vm_flags = vma->vm_flags;
120         pgprot_t vm_page_prot;
121
122         vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
123         if (vma_wants_writenotify(vma, vm_page_prot)) {
124                 vm_flags &= ~VM_SHARED;
125                 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
126         }
127         /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
128         WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
129 }
130
131 /*
132  * Requires inode->i_mapping->i_mmap_rwsem
133  */
134 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
135                 struct file *file, struct address_space *mapping)
136 {
137         if (vma->vm_flags & VM_DENYWRITE)
138                 atomic_inc(&file_inode(file)->i_writecount);
139         if (vma->vm_flags & VM_SHARED)
140                 mapping_unmap_writable(mapping);
141
142         flush_dcache_mmap_lock(mapping);
143         vma_interval_tree_remove(vma, &mapping->i_mmap);
144         flush_dcache_mmap_unlock(mapping);
145 }
146
147 /*
148  * Unlink a file-based vm structure from its interval tree, to hide
149  * vma from rmap and vmtruncate before freeing its page tables.
150  */
151 void unlink_file_vma(struct vm_area_struct *vma)
152 {
153         struct file *file = vma->vm_file;
154
155         if (file) {
156                 struct address_space *mapping = file->f_mapping;
157                 i_mmap_lock_write(mapping);
158                 __remove_shared_vm_struct(vma, file, mapping);
159                 i_mmap_unlock_write(mapping);
160         }
161 }
162
163 /*
164  * Close a vm structure and free it, returning the next.
165  */
166 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
167 {
168         struct vm_area_struct *next = vma->vm_next;
169
170         might_sleep();
171         if (vma->vm_ops && vma->vm_ops->close)
172                 vma->vm_ops->close(vma);
173         if (vma->vm_file)
174                 fput(vma->vm_file);
175         mpol_put(vma_policy(vma));
176         kmem_cache_free(vm_area_cachep, vma);
177         return next;
178 }
179
180 static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf);
181
182 SYSCALL_DEFINE1(brk, unsigned long, brk)
183 {
184         unsigned long retval;
185         unsigned long newbrk, oldbrk;
186         struct mm_struct *mm = current->mm;
187         struct vm_area_struct *next;
188         unsigned long min_brk;
189         bool populate;
190         LIST_HEAD(uf);
191
192         if (down_write_killable(&mm->mmap_sem))
193                 return -EINTR;
194
195 #ifdef CONFIG_COMPAT_BRK
196         /*
197          * CONFIG_COMPAT_BRK can still be overridden by setting
198          * randomize_va_space to 2, which will still cause mm->start_brk
199          * to be arbitrarily shifted
200          */
201         if (current->brk_randomized)
202                 min_brk = mm->start_brk;
203         else
204                 min_brk = mm->end_data;
205 #else
206         min_brk = mm->start_brk;
207 #endif
208         if (brk < min_brk)
209                 goto out;
210
211         /*
212          * Check against rlimit here. If this check is done later after the test
213          * of oldbrk with newbrk then it can escape the test and let the data
214          * segment grow beyond its set limit the in case where the limit is
215          * not page aligned -Ram Gupta
216          */
217         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
218                               mm->end_data, mm->start_data))
219                 goto out;
220
221         newbrk = PAGE_ALIGN(brk);
222         oldbrk = PAGE_ALIGN(mm->brk);
223         if (oldbrk == newbrk)
224                 goto set_brk;
225
226         /* Always allow shrinking brk. */
227         if (brk <= mm->brk) {
228                 if (!do_munmap(mm, newbrk, oldbrk-newbrk, &uf))
229                         goto set_brk;
230                 goto out;
231         }
232
233         /* Check against existing mmap mappings. */
234         next = find_vma(mm, oldbrk);
235         if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
236                 goto out;
237
238         /* Ok, looks good - let it rip. */
239         if (do_brk(oldbrk, newbrk-oldbrk, &uf) < 0)
240                 goto out;
241
242 set_brk:
243         mm->brk = brk;
244         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
245         up_write(&mm->mmap_sem);
246         userfaultfd_unmap_complete(mm, &uf);
247         if (populate)
248                 mm_populate(oldbrk, newbrk - oldbrk);
249         return brk;
250
251 out:
252         retval = mm->brk;
253         up_write(&mm->mmap_sem);
254         return retval;
255 }
256
257 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
258 {
259         unsigned long max, prev_end, subtree_gap;
260
261         /*
262          * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
263          * allow two stack_guard_gaps between them here, and when choosing
264          * an unmapped area; whereas when expanding we only require one.
265          * That's a little inconsistent, but keeps the code here simpler.
266          */
267         max = vm_start_gap(vma);
268         if (vma->vm_prev) {
269                 prev_end = vm_end_gap(vma->vm_prev);
270                 if (max > prev_end)
271                         max -= prev_end;
272                 else
273                         max = 0;
274         }
275         if (vma->vm_rb.rb_left) {
276                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
277                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
278                 if (subtree_gap > max)
279                         max = subtree_gap;
280         }
281         if (vma->vm_rb.rb_right) {
282                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
283                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
284                 if (subtree_gap > max)
285                         max = subtree_gap;
286         }
287         return max;
288 }
289
290 #ifdef CONFIG_DEBUG_VM_RB
291 static int browse_rb(struct mm_struct *mm)
292 {
293         struct rb_root *root = &mm->mm_rb;
294         int i = 0, j, bug = 0;
295         struct rb_node *nd, *pn = NULL;
296         unsigned long prev = 0, pend = 0;
297
298         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
299                 struct vm_area_struct *vma;
300                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
301                 if (vma->vm_start < prev) {
302                         pr_emerg("vm_start %lx < prev %lx\n",
303                                   vma->vm_start, prev);
304                         bug = 1;
305                 }
306                 if (vma->vm_start < pend) {
307                         pr_emerg("vm_start %lx < pend %lx\n",
308                                   vma->vm_start, pend);
309                         bug = 1;
310                 }
311                 if (vma->vm_start > vma->vm_end) {
312                         pr_emerg("vm_start %lx > vm_end %lx\n",
313                                   vma->vm_start, vma->vm_end);
314                         bug = 1;
315                 }
316                 spin_lock(&mm->page_table_lock);
317                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
318                         pr_emerg("free gap %lx, correct %lx\n",
319                                vma->rb_subtree_gap,
320                                vma_compute_subtree_gap(vma));
321                         bug = 1;
322                 }
323                 spin_unlock(&mm->page_table_lock);
324                 i++;
325                 pn = nd;
326                 prev = vma->vm_start;
327                 pend = vma->vm_end;
328         }
329         j = 0;
330         for (nd = pn; nd; nd = rb_prev(nd))
331                 j++;
332         if (i != j) {
333                 pr_emerg("backwards %d, forwards %d\n", j, i);
334                 bug = 1;
335         }
336         return bug ? -1 : i;
337 }
338
339 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
340 {
341         struct rb_node *nd;
342
343         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
344                 struct vm_area_struct *vma;
345                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
346                 VM_BUG_ON_VMA(vma != ignore &&
347                         vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
348                         vma);
349         }
350 }
351
352 static void validate_mm(struct mm_struct *mm)
353 {
354         int bug = 0;
355         int i = 0;
356         unsigned long highest_address = 0;
357         struct vm_area_struct *vma = mm->mmap;
358
359         while (vma) {
360                 struct anon_vma *anon_vma = vma->anon_vma;
361                 struct anon_vma_chain *avc;
362
363                 if (anon_vma) {
364                         anon_vma_lock_read(anon_vma);
365                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
366                                 anon_vma_interval_tree_verify(avc);
367                         anon_vma_unlock_read(anon_vma);
368                 }
369
370                 highest_address = vm_end_gap(vma);
371                 vma = vma->vm_next;
372                 i++;
373         }
374         if (i != mm->map_count) {
375                 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
376                 bug = 1;
377         }
378         if (highest_address != mm->highest_vm_end) {
379                 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
380                           mm->highest_vm_end, highest_address);
381                 bug = 1;
382         }
383         i = browse_rb(mm);
384         if (i != mm->map_count) {
385                 if (i != -1)
386                         pr_emerg("map_count %d rb %d\n", mm->map_count, i);
387                 bug = 1;
388         }
389         VM_BUG_ON_MM(bug, mm);
390 }
391 #else
392 #define validate_mm_rb(root, ignore) do { } while (0)
393 #define validate_mm(mm) do { } while (0)
394 #endif
395
396 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
397                      unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
398
399 /*
400  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
401  * vma->vm_prev->vm_end values changed, without modifying the vma's position
402  * in the rbtree.
403  */
404 static void vma_gap_update(struct vm_area_struct *vma)
405 {
406         /*
407          * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
408          * function that does exacltly what we want.
409          */
410         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
411 }
412
413 static inline void vma_rb_insert(struct vm_area_struct *vma,
414                                  struct rb_root *root)
415 {
416         /* All rb_subtree_gap values must be consistent prior to insertion */
417         validate_mm_rb(root, NULL);
418
419         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
420 }
421
422 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
423 {
424         /*
425          * Note rb_erase_augmented is a fairly large inline function,
426          * so make sure we instantiate it only once with our desired
427          * augmented rbtree callbacks.
428          */
429         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
430 }
431
432 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
433                                                 struct rb_root *root,
434                                                 struct vm_area_struct *ignore)
435 {
436         /*
437          * All rb_subtree_gap values must be consistent prior to erase,
438          * with the possible exception of the "next" vma being erased if
439          * next->vm_start was reduced.
440          */
441         validate_mm_rb(root, ignore);
442
443         __vma_rb_erase(vma, root);
444 }
445
446 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
447                                          struct rb_root *root)
448 {
449         /*
450          * All rb_subtree_gap values must be consistent prior to erase,
451          * with the possible exception of the vma being erased.
452          */
453         validate_mm_rb(root, vma);
454
455         __vma_rb_erase(vma, root);
456 }
457
458 /*
459  * vma has some anon_vma assigned, and is already inserted on that
460  * anon_vma's interval trees.
461  *
462  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
463  * vma must be removed from the anon_vma's interval trees using
464  * anon_vma_interval_tree_pre_update_vma().
465  *
466  * After the update, the vma will be reinserted using
467  * anon_vma_interval_tree_post_update_vma().
468  *
469  * The entire update must be protected by exclusive mmap_sem and by
470  * the root anon_vma's mutex.
471  */
472 static inline void
473 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
474 {
475         struct anon_vma_chain *avc;
476
477         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
478                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
479 }
480
481 static inline void
482 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
483 {
484         struct anon_vma_chain *avc;
485
486         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
487                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
488 }
489
490 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
491                 unsigned long end, struct vm_area_struct **pprev,
492                 struct rb_node ***rb_link, struct rb_node **rb_parent)
493 {
494         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
495
496         __rb_link = &mm->mm_rb.rb_node;
497         rb_prev = __rb_parent = NULL;
498
499         while (*__rb_link) {
500                 struct vm_area_struct *vma_tmp;
501
502                 __rb_parent = *__rb_link;
503                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
504
505                 if (vma_tmp->vm_end > addr) {
506                         /* Fail if an existing vma overlaps the area */
507                         if (vma_tmp->vm_start < end)
508                                 return -ENOMEM;
509                         __rb_link = &__rb_parent->rb_left;
510                 } else {
511                         rb_prev = __rb_parent;
512                         __rb_link = &__rb_parent->rb_right;
513                 }
514         }
515
516         *pprev = NULL;
517         if (rb_prev)
518                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
519         *rb_link = __rb_link;
520         *rb_parent = __rb_parent;
521         return 0;
522 }
523
524 static unsigned long count_vma_pages_range(struct mm_struct *mm,
525                 unsigned long addr, unsigned long end)
526 {
527         unsigned long nr_pages = 0;
528         struct vm_area_struct *vma;
529
530         /* Find first overlaping mapping */
531         vma = find_vma_intersection(mm, addr, end);
532         if (!vma)
533                 return 0;
534
535         nr_pages = (min(end, vma->vm_end) -
536                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
537
538         /* Iterate over the rest of the overlaps */
539         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
540                 unsigned long overlap_len;
541
542                 if (vma->vm_start > end)
543                         break;
544
545                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
546                 nr_pages += overlap_len >> PAGE_SHIFT;
547         }
548
549         return nr_pages;
550 }
551
552 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
553                 struct rb_node **rb_link, struct rb_node *rb_parent)
554 {
555         /* Update tracking information for the gap following the new vma. */
556         if (vma->vm_next)
557                 vma_gap_update(vma->vm_next);
558         else
559                 mm->highest_vm_end = vm_end_gap(vma);
560
561         /*
562          * vma->vm_prev wasn't known when we followed the rbtree to find the
563          * correct insertion point for that vma. As a result, we could not
564          * update the vma vm_rb parents rb_subtree_gap values on the way down.
565          * So, we first insert the vma with a zero rb_subtree_gap value
566          * (to be consistent with what we did on the way down), and then
567          * immediately update the gap to the correct value. Finally we
568          * rebalance the rbtree after all augmented values have been set.
569          */
570         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
571         vma->rb_subtree_gap = 0;
572         vma_gap_update(vma);
573         vma_rb_insert(vma, &mm->mm_rb);
574 }
575
576 static void __vma_link_file(struct vm_area_struct *vma)
577 {
578         struct file *file;
579
580         file = vma->vm_file;
581         if (file) {
582                 struct address_space *mapping = file->f_mapping;
583
584                 if (vma->vm_flags & VM_DENYWRITE)
585                         atomic_dec(&file_inode(file)->i_writecount);
586                 if (vma->vm_flags & VM_SHARED)
587                         atomic_inc(&mapping->i_mmap_writable);
588
589                 flush_dcache_mmap_lock(mapping);
590                 vma_interval_tree_insert(vma, &mapping->i_mmap);
591                 flush_dcache_mmap_unlock(mapping);
592         }
593 }
594
595 static void
596 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
597         struct vm_area_struct *prev, struct rb_node **rb_link,
598         struct rb_node *rb_parent)
599 {
600         __vma_link_list(mm, vma, prev, rb_parent);
601         __vma_link_rb(mm, vma, rb_link, rb_parent);
602 }
603
604 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
605                         struct vm_area_struct *prev, struct rb_node **rb_link,
606                         struct rb_node *rb_parent)
607 {
608         struct address_space *mapping = NULL;
609
610         if (vma->vm_file) {
611                 mapping = vma->vm_file->f_mapping;
612                 i_mmap_lock_write(mapping);
613         }
614
615         __vma_link(mm, vma, prev, rb_link, rb_parent);
616         __vma_link_file(vma);
617
618         if (mapping)
619                 i_mmap_unlock_write(mapping);
620
621         mm->map_count++;
622         validate_mm(mm);
623 }
624
625 /*
626  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
627  * mm's list and rbtree.  It has already been inserted into the interval tree.
628  */
629 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
630 {
631         struct vm_area_struct *prev;
632         struct rb_node **rb_link, *rb_parent;
633
634         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
635                            &prev, &rb_link, &rb_parent))
636                 BUG();
637         __vma_link(mm, vma, prev, rb_link, rb_parent);
638         mm->map_count++;
639 }
640
641 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
642                                                 struct vm_area_struct *vma,
643                                                 struct vm_area_struct *prev,
644                                                 bool has_prev,
645                                                 struct vm_area_struct *ignore)
646 {
647         struct vm_area_struct *next;
648
649         vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
650         next = vma->vm_next;
651         if (has_prev)
652                 prev->vm_next = next;
653         else {
654                 prev = vma->vm_prev;
655                 if (prev)
656                         prev->vm_next = next;
657                 else
658                         mm->mmap = next;
659         }
660         if (next)
661                 next->vm_prev = prev;
662
663         /* Kill the cache */
664         vmacache_invalidate(mm);
665 }
666
667 static inline void __vma_unlink_prev(struct mm_struct *mm,
668                                      struct vm_area_struct *vma,
669                                      struct vm_area_struct *prev)
670 {
671         __vma_unlink_common(mm, vma, prev, true, vma);
672 }
673
674 /*
675  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
676  * is already present in an i_mmap tree without adjusting the tree.
677  * The following helper function should be used when such adjustments
678  * are necessary.  The "insert" vma (if any) is to be inserted
679  * before we drop the necessary locks.
680  */
681 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
682         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
683         struct vm_area_struct *expand)
684 {
685         struct mm_struct *mm = vma->vm_mm;
686         struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
687         struct address_space *mapping = NULL;
688         struct rb_root_cached *root = NULL;
689         struct anon_vma *anon_vma = NULL;
690         struct file *file = vma->vm_file;
691         bool start_changed = false, end_changed = false;
692         long adjust_next = 0;
693         int remove_next = 0;
694
695         if (next && !insert) {
696                 struct vm_area_struct *exporter = NULL, *importer = NULL;
697
698                 if (end >= next->vm_end) {
699                         /*
700                          * vma expands, overlapping all the next, and
701                          * perhaps the one after too (mprotect case 6).
702                          * The only other cases that gets here are
703                          * case 1, case 7 and case 8.
704                          */
705                         if (next == expand) {
706                                 /*
707                                  * The only case where we don't expand "vma"
708                                  * and we expand "next" instead is case 8.
709                                  */
710                                 VM_WARN_ON(end != next->vm_end);
711                                 /*
712                                  * remove_next == 3 means we're
713                                  * removing "vma" and that to do so we
714                                  * swapped "vma" and "next".
715                                  */
716                                 remove_next = 3;
717                                 VM_WARN_ON(file != next->vm_file);
718                                 swap(vma, next);
719                         } else {
720                                 VM_WARN_ON(expand != vma);
721                                 /*
722                                  * case 1, 6, 7, remove_next == 2 is case 6,
723                                  * remove_next == 1 is case 1 or 7.
724                                  */
725                                 remove_next = 1 + (end > next->vm_end);
726                                 VM_WARN_ON(remove_next == 2 &&
727                                            end != next->vm_next->vm_end);
728                                 VM_WARN_ON(remove_next == 1 &&
729                                            end != next->vm_end);
730                                 /* trim end to next, for case 6 first pass */
731                                 end = next->vm_end;
732                         }
733
734                         exporter = next;
735                         importer = vma;
736
737                         /*
738                          * If next doesn't have anon_vma, import from vma after
739                          * next, if the vma overlaps with it.
740                          */
741                         if (remove_next == 2 && !next->anon_vma)
742                                 exporter = next->vm_next;
743
744                 } else if (end > next->vm_start) {
745                         /*
746                          * vma expands, overlapping part of the next:
747                          * mprotect case 5 shifting the boundary up.
748                          */
749                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
750                         exporter = next;
751                         importer = vma;
752                         VM_WARN_ON(expand != importer);
753                 } else if (end < vma->vm_end) {
754                         /*
755                          * vma shrinks, and !insert tells it's not
756                          * split_vma inserting another: so it must be
757                          * mprotect case 4 shifting the boundary down.
758                          */
759                         adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
760                         exporter = vma;
761                         importer = next;
762                         VM_WARN_ON(expand != importer);
763                 }
764
765                 /*
766                  * Easily overlooked: when mprotect shifts the boundary,
767                  * make sure the expanding vma has anon_vma set if the
768                  * shrinking vma had, to cover any anon pages imported.
769                  */
770                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
771                         int error;
772
773                         importer->anon_vma = exporter->anon_vma;
774                         error = anon_vma_clone(importer, exporter);
775                         if (error)
776                                 return error;
777                 }
778         }
779 again:
780         vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
781
782         if (file) {
783                 mapping = file->f_mapping;
784                 root = &mapping->i_mmap;
785                 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
786
787                 if (adjust_next)
788                         uprobe_munmap(next, next->vm_start, next->vm_end);
789
790                 i_mmap_lock_write(mapping);
791                 if (insert) {
792                         /*
793                          * Put into interval tree now, so instantiated pages
794                          * are visible to arm/parisc __flush_dcache_page
795                          * throughout; but we cannot insert into address
796                          * space until vma start or end is updated.
797                          */
798                         __vma_link_file(insert);
799                 }
800         }
801
802         anon_vma = vma->anon_vma;
803         if (!anon_vma && adjust_next)
804                 anon_vma = next->anon_vma;
805         if (anon_vma) {
806                 VM_WARN_ON(adjust_next && next->anon_vma &&
807                            anon_vma != next->anon_vma);
808                 anon_vma_lock_write(anon_vma);
809                 anon_vma_interval_tree_pre_update_vma(vma);
810                 if (adjust_next)
811                         anon_vma_interval_tree_pre_update_vma(next);
812         }
813
814         if (root) {
815                 flush_dcache_mmap_lock(mapping);
816                 vma_interval_tree_remove(vma, root);
817                 if (adjust_next)
818                         vma_interval_tree_remove(next, root);
819         }
820
821         if (start != vma->vm_start) {
822                 vma->vm_start = start;
823                 start_changed = true;
824         }
825         if (end != vma->vm_end) {
826                 vma->vm_end = end;
827                 end_changed = true;
828         }
829         vma->vm_pgoff = pgoff;
830         if (adjust_next) {
831                 next->vm_start += adjust_next << PAGE_SHIFT;
832                 next->vm_pgoff += adjust_next;
833         }
834
835         if (root) {
836                 if (adjust_next)
837                         vma_interval_tree_insert(next, root);
838                 vma_interval_tree_insert(vma, root);
839                 flush_dcache_mmap_unlock(mapping);
840         }
841
842         if (remove_next) {
843                 /*
844                  * vma_merge has merged next into vma, and needs
845                  * us to remove next before dropping the locks.
846                  */
847                 if (remove_next != 3)
848                         __vma_unlink_prev(mm, next, vma);
849                 else
850                         /*
851                          * vma is not before next if they've been
852                          * swapped.
853                          *
854                          * pre-swap() next->vm_start was reduced so
855                          * tell validate_mm_rb to ignore pre-swap()
856                          * "next" (which is stored in post-swap()
857                          * "vma").
858                          */
859                         __vma_unlink_common(mm, next, NULL, false, vma);
860                 if (file)
861                         __remove_shared_vm_struct(next, file, mapping);
862         } else if (insert) {
863                 /*
864                  * split_vma has split insert from vma, and needs
865                  * us to insert it before dropping the locks
866                  * (it may either follow vma or precede it).
867                  */
868                 __insert_vm_struct(mm, insert);
869         } else {
870                 if (start_changed)
871                         vma_gap_update(vma);
872                 if (end_changed) {
873                         if (!next)
874                                 mm->highest_vm_end = vm_end_gap(vma);
875                         else if (!adjust_next)
876                                 vma_gap_update(next);
877                 }
878         }
879
880         if (anon_vma) {
881                 anon_vma_interval_tree_post_update_vma(vma);
882                 if (adjust_next)
883                         anon_vma_interval_tree_post_update_vma(next);
884                 anon_vma_unlock_write(anon_vma);
885         }
886         if (mapping)
887                 i_mmap_unlock_write(mapping);
888
889         if (root) {
890                 uprobe_mmap(vma);
891
892                 if (adjust_next)
893                         uprobe_mmap(next);
894         }
895
896         if (remove_next) {
897                 if (file) {
898                         uprobe_munmap(next, next->vm_start, next->vm_end);
899                         fput(file);
900                 }
901                 if (next->anon_vma)
902                         anon_vma_merge(vma, next);
903                 mm->map_count--;
904                 mpol_put(vma_policy(next));
905                 kmem_cache_free(vm_area_cachep, next);
906                 /*
907                  * In mprotect's case 6 (see comments on vma_merge),
908                  * we must remove another next too. It would clutter
909                  * up the code too much to do both in one go.
910                  */
911                 if (remove_next != 3) {
912                         /*
913                          * If "next" was removed and vma->vm_end was
914                          * expanded (up) over it, in turn
915                          * "next->vm_prev->vm_end" changed and the
916                          * "vma->vm_next" gap must be updated.
917                          */
918                         next = vma->vm_next;
919                 } else {
920                         /*
921                          * For the scope of the comment "next" and
922                          * "vma" considered pre-swap(): if "vma" was
923                          * removed, next->vm_start was expanded (down)
924                          * over it and the "next" gap must be updated.
925                          * Because of the swap() the post-swap() "vma"
926                          * actually points to pre-swap() "next"
927                          * (post-swap() "next" as opposed is now a
928                          * dangling pointer).
929                          */
930                         next = vma;
931                 }
932                 if (remove_next == 2) {
933                         remove_next = 1;
934                         end = next->vm_end;
935                         goto again;
936                 }
937                 else if (next)
938                         vma_gap_update(next);
939                 else {
940                         /*
941                          * If remove_next == 2 we obviously can't
942                          * reach this path.
943                          *
944                          * If remove_next == 3 we can't reach this
945                          * path because pre-swap() next is always not
946                          * NULL. pre-swap() "next" is not being
947                          * removed and its next->vm_end is not altered
948                          * (and furthermore "end" already matches
949                          * next->vm_end in remove_next == 3).
950                          *
951                          * We reach this only in the remove_next == 1
952                          * case if the "next" vma that was removed was
953                          * the highest vma of the mm. However in such
954                          * case next->vm_end == "end" and the extended
955                          * "vma" has vma->vm_end == next->vm_end so
956                          * mm->highest_vm_end doesn't need any update
957                          * in remove_next == 1 case.
958                          */
959                         VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
960                 }
961         }
962         if (insert && file)
963                 uprobe_mmap(insert);
964
965         validate_mm(mm);
966
967         return 0;
968 }
969
970 /*
971  * If the vma has a ->close operation then the driver probably needs to release
972  * per-vma resources, so we don't attempt to merge those.
973  */
974 static inline int is_mergeable_vma(struct vm_area_struct *vma,
975                                 struct file *file, unsigned long vm_flags,
976                                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
977 {
978         /*
979          * VM_SOFTDIRTY should not prevent from VMA merging, if we
980          * match the flags but dirty bit -- the caller should mark
981          * merged VMA as dirty. If dirty bit won't be excluded from
982          * comparison, we increase pressue on the memory system forcing
983          * the kernel to generate new VMAs when old one could be
984          * extended instead.
985          */
986         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
987                 return 0;
988         if (vma->vm_file != file)
989                 return 0;
990         if (vma->vm_ops && vma->vm_ops->close)
991                 return 0;
992         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
993                 return 0;
994         return 1;
995 }
996
997 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
998                                         struct anon_vma *anon_vma2,
999                                         struct vm_area_struct *vma)
1000 {
1001         /*
1002          * The list_is_singular() test is to avoid merging VMA cloned from
1003          * parents. This can improve scalability caused by anon_vma lock.
1004          */
1005         if ((!anon_vma1 || !anon_vma2) && (!vma ||
1006                 list_is_singular(&vma->anon_vma_chain)))
1007                 return 1;
1008         return anon_vma1 == anon_vma2;
1009 }
1010
1011 /*
1012  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1013  * in front of (at a lower virtual address and file offset than) the vma.
1014  *
1015  * We cannot merge two vmas if they have differently assigned (non-NULL)
1016  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1017  *
1018  * We don't check here for the merged mmap wrapping around the end of pagecache
1019  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1020  * wrap, nor mmaps which cover the final page at index -1UL.
1021  */
1022 static int
1023 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1024                      struct anon_vma *anon_vma, struct file *file,
1025                      pgoff_t vm_pgoff,
1026                      struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1027 {
1028         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1029             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1030                 if (vma->vm_pgoff == vm_pgoff)
1031                         return 1;
1032         }
1033         return 0;
1034 }
1035
1036 /*
1037  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1038  * beyond (at a higher virtual address and file offset than) the vma.
1039  *
1040  * We cannot merge two vmas if they have differently assigned (non-NULL)
1041  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1042  */
1043 static int
1044 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1045                     struct anon_vma *anon_vma, struct file *file,
1046                     pgoff_t vm_pgoff,
1047                     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1048 {
1049         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1050             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1051                 pgoff_t vm_pglen;
1052                 vm_pglen = vma_pages(vma);
1053                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1054                         return 1;
1055         }
1056         return 0;
1057 }
1058
1059 /*
1060  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1061  * whether that can be merged with its predecessor or its successor.
1062  * Or both (it neatly fills a hole).
1063  *
1064  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1065  * certain not to be mapped by the time vma_merge is called; but when
1066  * called for mprotect, it is certain to be already mapped (either at
1067  * an offset within prev, or at the start of next), and the flags of
1068  * this area are about to be changed to vm_flags - and the no-change
1069  * case has already been eliminated.
1070  *
1071  * The following mprotect cases have to be considered, where AAAA is
1072  * the area passed down from mprotect_fixup, never extending beyond one
1073  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1074  *
1075  *     AAAA             AAAA                AAAA          AAAA
1076  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
1077  *    cannot merge    might become    might become    might become
1078  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
1079  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
1080  *    mremap move:                                    PPPPXXXXXXXX 8
1081  *        AAAA
1082  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
1083  *    might become    case 1 below    case 2 below    case 3 below
1084  *
1085  * It is important for case 8 that the the vma NNNN overlapping the
1086  * region AAAA is never going to extended over XXXX. Instead XXXX must
1087  * be extended in region AAAA and NNNN must be removed. This way in
1088  * all cases where vma_merge succeeds, the moment vma_adjust drops the
1089  * rmap_locks, the properties of the merged vma will be already
1090  * correct for the whole merged range. Some of those properties like
1091  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1092  * be correct for the whole merged range immediately after the
1093  * rmap_locks are released. Otherwise if XXXX would be removed and
1094  * NNNN would be extended over the XXXX range, remove_migration_ptes
1095  * or other rmap walkers (if working on addresses beyond the "end"
1096  * parameter) may establish ptes with the wrong permissions of NNNN
1097  * instead of the right permissions of XXXX.
1098  */
1099 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1100                         struct vm_area_struct *prev, unsigned long addr,
1101                         unsigned long end, unsigned long vm_flags,
1102                         struct anon_vma *anon_vma, struct file *file,
1103                         pgoff_t pgoff, struct mempolicy *policy,
1104                         struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1105 {
1106         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1107         struct vm_area_struct *area, *next;
1108         int err;
1109
1110         /*
1111          * We later require that vma->vm_flags == vm_flags,
1112          * so this tests vma->vm_flags & VM_SPECIAL, too.
1113          */
1114         if (vm_flags & VM_SPECIAL)
1115                 return NULL;
1116
1117         if (prev)
1118                 next = prev->vm_next;
1119         else
1120                 next = mm->mmap;
1121         area = next;
1122         if (area && area->vm_end == end)                /* cases 6, 7, 8 */
1123                 next = next->vm_next;
1124
1125         /* verify some invariant that must be enforced by the caller */
1126         VM_WARN_ON(prev && addr <= prev->vm_start);
1127         VM_WARN_ON(area && end > area->vm_end);
1128         VM_WARN_ON(addr >= end);
1129
1130         /*
1131          * Can it merge with the predecessor?
1132          */
1133         if (prev && prev->vm_end == addr &&
1134                         mpol_equal(vma_policy(prev), policy) &&
1135                         can_vma_merge_after(prev, vm_flags,
1136                                             anon_vma, file, pgoff,
1137                                             vm_userfaultfd_ctx)) {
1138                 /*
1139                  * OK, it can.  Can we now merge in the successor as well?
1140                  */
1141                 if (next && end == next->vm_start &&
1142                                 mpol_equal(policy, vma_policy(next)) &&
1143                                 can_vma_merge_before(next, vm_flags,
1144                                                      anon_vma, file,
1145                                                      pgoff+pglen,
1146                                                      vm_userfaultfd_ctx) &&
1147                                 is_mergeable_anon_vma(prev->anon_vma,
1148                                                       next->anon_vma, NULL)) {
1149                                                         /* cases 1, 6 */
1150                         err = __vma_adjust(prev, prev->vm_start,
1151                                          next->vm_end, prev->vm_pgoff, NULL,
1152                                          prev);
1153                 } else                                  /* cases 2, 5, 7 */
1154                         err = __vma_adjust(prev, prev->vm_start,
1155                                          end, prev->vm_pgoff, NULL, prev);
1156                 if (err)
1157                         return NULL;
1158                 khugepaged_enter_vma_merge(prev, vm_flags);
1159                 return prev;
1160         }
1161
1162         /*
1163          * Can this new request be merged in front of next?
1164          */
1165         if (next && end == next->vm_start &&
1166                         mpol_equal(policy, vma_policy(next)) &&
1167                         can_vma_merge_before(next, vm_flags,
1168                                              anon_vma, file, pgoff+pglen,
1169                                              vm_userfaultfd_ctx)) {
1170                 if (prev && addr < prev->vm_end)        /* case 4 */
1171                         err = __vma_adjust(prev, prev->vm_start,
1172                                          addr, prev->vm_pgoff, NULL, next);
1173                 else {                                  /* cases 3, 8 */
1174                         err = __vma_adjust(area, addr, next->vm_end,
1175                                          next->vm_pgoff - pglen, NULL, next);
1176                         /*
1177                          * In case 3 area is already equal to next and
1178                          * this is a noop, but in case 8 "area" has
1179                          * been removed and next was expanded over it.
1180                          */
1181                         area = next;
1182                 }
1183                 if (err)
1184                         return NULL;
1185                 khugepaged_enter_vma_merge(area, vm_flags);
1186                 return area;
1187         }
1188
1189         return NULL;
1190 }
1191
1192 /*
1193  * Rough compatbility check to quickly see if it's even worth looking
1194  * at sharing an anon_vma.
1195  *
1196  * They need to have the same vm_file, and the flags can only differ
1197  * in things that mprotect may change.
1198  *
1199  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1200  * we can merge the two vma's. For example, we refuse to merge a vma if
1201  * there is a vm_ops->close() function, because that indicates that the
1202  * driver is doing some kind of reference counting. But that doesn't
1203  * really matter for the anon_vma sharing case.
1204  */
1205 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1206 {
1207         return a->vm_end == b->vm_start &&
1208                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1209                 a->vm_file == b->vm_file &&
1210                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1211                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1212 }
1213
1214 /*
1215  * Do some basic sanity checking to see if we can re-use the anon_vma
1216  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1217  * the same as 'old', the other will be the new one that is trying
1218  * to share the anon_vma.
1219  *
1220  * NOTE! This runs with mm_sem held for reading, so it is possible that
1221  * the anon_vma of 'old' is concurrently in the process of being set up
1222  * by another page fault trying to merge _that_. But that's ok: if it
1223  * is being set up, that automatically means that it will be a singleton
1224  * acceptable for merging, so we can do all of this optimistically. But
1225  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1226  *
1227  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1228  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1229  * is to return an anon_vma that is "complex" due to having gone through
1230  * a fork).
1231  *
1232  * We also make sure that the two vma's are compatible (adjacent,
1233  * and with the same memory policies). That's all stable, even with just
1234  * a read lock on the mm_sem.
1235  */
1236 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1237 {
1238         if (anon_vma_compatible(a, b)) {
1239                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1240
1241                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1242                         return anon_vma;
1243         }
1244         return NULL;
1245 }
1246
1247 /*
1248  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1249  * neighbouring vmas for a suitable anon_vma, before it goes off
1250  * to allocate a new anon_vma.  It checks because a repetitive
1251  * sequence of mprotects and faults may otherwise lead to distinct
1252  * anon_vmas being allocated, preventing vma merge in subsequent
1253  * mprotect.
1254  */
1255 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1256 {
1257         struct anon_vma *anon_vma;
1258         struct vm_area_struct *near;
1259
1260         near = vma->vm_next;
1261         if (!near)
1262                 goto try_prev;
1263
1264         anon_vma = reusable_anon_vma(near, vma, near);
1265         if (anon_vma)
1266                 return anon_vma;
1267 try_prev:
1268         near = vma->vm_prev;
1269         if (!near)
1270                 goto none;
1271
1272         anon_vma = reusable_anon_vma(near, near, vma);
1273         if (anon_vma)
1274                 return anon_vma;
1275 none:
1276         /*
1277          * There's no absolute need to look only at touching neighbours:
1278          * we could search further afield for "compatible" anon_vmas.
1279          * But it would probably just be a waste of time searching,
1280          * or lead to too many vmas hanging off the same anon_vma.
1281          * We're trying to allow mprotect remerging later on,
1282          * not trying to minimize memory used for anon_vmas.
1283          */
1284         return NULL;
1285 }
1286
1287 /*
1288  * If a hint addr is less than mmap_min_addr change hint to be as
1289  * low as possible but still greater than mmap_min_addr
1290  */
1291 static inline unsigned long round_hint_to_min(unsigned long hint)
1292 {
1293         hint &= PAGE_MASK;
1294         if (((void *)hint != NULL) &&
1295             (hint < mmap_min_addr))
1296                 return PAGE_ALIGN(mmap_min_addr);
1297         return hint;
1298 }
1299
1300 static inline int mlock_future_check(struct mm_struct *mm,
1301                                      unsigned long flags,
1302                                      unsigned long len)
1303 {
1304         unsigned long locked, lock_limit;
1305
1306         /*  mlock MCL_FUTURE? */
1307         if (flags & VM_LOCKED) {
1308                 locked = len >> PAGE_SHIFT;
1309                 locked += mm->locked_vm;
1310                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1311                 lock_limit >>= PAGE_SHIFT;
1312                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1313                         return -EAGAIN;
1314         }
1315         return 0;
1316 }
1317
1318 /*
1319  * The caller must hold down_write(&current->mm->mmap_sem).
1320  */
1321 unsigned long do_mmap(struct file *file, unsigned long addr,
1322                         unsigned long len, unsigned long prot,
1323                         unsigned long flags, vm_flags_t vm_flags,
1324                         unsigned long pgoff, unsigned long *populate,
1325                         struct list_head *uf)
1326 {
1327         struct mm_struct *mm = current->mm;
1328         int pkey = 0;
1329
1330         *populate = 0;
1331
1332         if (!len)
1333                 return -EINVAL;
1334
1335         /*
1336          * Does the application expect PROT_READ to imply PROT_EXEC?
1337          *
1338          * (the exception is when the underlying filesystem is noexec
1339          *  mounted, in which case we dont add PROT_EXEC.)
1340          */
1341         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1342                 if (!(file && path_noexec(&file->f_path)))
1343                         prot |= PROT_EXEC;
1344
1345         if (!(flags & MAP_FIXED))
1346                 addr = round_hint_to_min(addr);
1347
1348         /* Careful about overflows.. */
1349         len = PAGE_ALIGN(len);
1350         if (!len)
1351                 return -ENOMEM;
1352
1353         /* offset overflow? */
1354         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1355                 return -EOVERFLOW;
1356
1357         /* Too many mappings? */
1358         if (mm->map_count > sysctl_max_map_count)
1359                 return -ENOMEM;
1360
1361         /* Obtain the address to map to. we verify (or select) it and ensure
1362          * that it represents a valid section of the address space.
1363          */
1364         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1365         if (offset_in_page(addr))
1366                 return addr;
1367
1368         if (prot == PROT_EXEC) {
1369                 pkey = execute_only_pkey(mm);
1370                 if (pkey < 0)
1371                         pkey = 0;
1372         }
1373
1374         /* Do simple checking here so the lower-level routines won't have
1375          * to. we assume access permissions have been handled by the open
1376          * of the memory object, so we don't do any here.
1377          */
1378         vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1379                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1380
1381         if (flags & MAP_LOCKED)
1382                 if (!can_do_mlock())
1383                         return -EPERM;
1384
1385         if (mlock_future_check(mm, vm_flags, len))
1386                 return -EAGAIN;
1387
1388         if (file) {
1389                 struct inode *inode = file_inode(file);
1390                 unsigned long flags_mask;
1391
1392                 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1393
1394                 switch (flags & MAP_TYPE) {
1395                 case MAP_SHARED:
1396                         /*
1397                          * Force use of MAP_SHARED_VALIDATE with non-legacy
1398                          * flags. E.g. MAP_SYNC is dangerous to use with
1399                          * MAP_SHARED as you don't know which consistency model
1400                          * you will get. We silently ignore unsupported flags
1401                          * with MAP_SHARED to preserve backward compatibility.
1402                          */
1403                         flags &= LEGACY_MAP_MASK;
1404                         /* fall through */
1405                 case MAP_SHARED_VALIDATE:
1406                         if (flags & ~flags_mask)
1407                                 return -EOPNOTSUPP;
1408                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1409                                 return -EACCES;
1410
1411                         /*
1412                          * Make sure we don't allow writing to an append-only
1413                          * file..
1414                          */
1415                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1416                                 return -EACCES;
1417
1418                         /*
1419                          * Make sure there are no mandatory locks on the file.
1420                          */
1421                         if (locks_verify_locked(file))
1422                                 return -EAGAIN;
1423
1424                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1425                         if (!(file->f_mode & FMODE_WRITE))
1426                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1427
1428                         /* fall through */
1429                 case MAP_PRIVATE:
1430                         if (!(file->f_mode & FMODE_READ))
1431                                 return -EACCES;
1432                         if (path_noexec(&file->f_path)) {
1433                                 if (vm_flags & VM_EXEC)
1434                                         return -EPERM;
1435                                 vm_flags &= ~VM_MAYEXEC;
1436                         }
1437
1438                         if (!file->f_op->mmap)
1439                                 return -ENODEV;
1440                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1441                                 return -EINVAL;
1442                         break;
1443
1444                 default:
1445                         return -EINVAL;
1446                 }
1447         } else {
1448                 switch (flags & MAP_TYPE) {
1449                 case MAP_SHARED:
1450                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1451                                 return -EINVAL;
1452                         /*
1453                          * Ignore pgoff.
1454                          */
1455                         pgoff = 0;
1456                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1457                         break;
1458                 case MAP_PRIVATE:
1459                         /*
1460                          * Set pgoff according to addr for anon_vma.
1461                          */
1462                         pgoff = addr >> PAGE_SHIFT;
1463                         break;
1464                 default:
1465                         return -EINVAL;
1466                 }
1467         }
1468
1469         /*
1470          * Set 'VM_NORESERVE' if we should not account for the
1471          * memory use of this mapping.
1472          */
1473         if (flags & MAP_NORESERVE) {
1474                 /* We honor MAP_NORESERVE if allowed to overcommit */
1475                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1476                         vm_flags |= VM_NORESERVE;
1477
1478                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1479                 if (file && is_file_hugepages(file))
1480                         vm_flags |= VM_NORESERVE;
1481         }
1482
1483         addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1484         if (!IS_ERR_VALUE(addr) &&
1485             ((vm_flags & VM_LOCKED) ||
1486              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1487                 *populate = len;
1488         return addr;
1489 }
1490
1491 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1492                               unsigned long prot, unsigned long flags,
1493                               unsigned long fd, unsigned long pgoff)
1494 {
1495         struct file *file = NULL;
1496         unsigned long retval;
1497
1498         if (!(flags & MAP_ANONYMOUS)) {
1499                 audit_mmap_fd(fd, flags);
1500                 file = fget(fd);
1501                 if (!file)
1502                         return -EBADF;
1503                 if (is_file_hugepages(file))
1504                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1505                 retval = -EINVAL;
1506                 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1507                         goto out_fput;
1508         } else if (flags & MAP_HUGETLB) {
1509                 struct user_struct *user = NULL;
1510                 struct hstate *hs;
1511
1512                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1513                 if (!hs)
1514                         return -EINVAL;
1515
1516                 len = ALIGN(len, huge_page_size(hs));
1517                 /*
1518                  * VM_NORESERVE is used because the reservations will be
1519                  * taken when vm_ops->mmap() is called
1520                  * A dummy user value is used because we are not locking
1521                  * memory so no accounting is necessary
1522                  */
1523                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1524                                 VM_NORESERVE,
1525                                 &user, HUGETLB_ANONHUGE_INODE,
1526                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1527                 if (IS_ERR(file))
1528                         return PTR_ERR(file);
1529         }
1530
1531         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1532
1533         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1534 out_fput:
1535         if (file)
1536                 fput(file);
1537         return retval;
1538 }
1539
1540 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1541                 unsigned long, prot, unsigned long, flags,
1542                 unsigned long, fd, unsigned long, pgoff)
1543 {
1544         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1545 }
1546
1547 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1548 struct mmap_arg_struct {
1549         unsigned long addr;
1550         unsigned long len;
1551         unsigned long prot;
1552         unsigned long flags;
1553         unsigned long fd;
1554         unsigned long offset;
1555 };
1556
1557 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1558 {
1559         struct mmap_arg_struct a;
1560
1561         if (copy_from_user(&a, arg, sizeof(a)))
1562                 return -EFAULT;
1563         if (offset_in_page(a.offset))
1564                 return -EINVAL;
1565
1566         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1567                                a.offset >> PAGE_SHIFT);
1568 }
1569 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1570
1571 /*
1572  * Some shared mappigns will want the pages marked read-only
1573  * to track write events. If so, we'll downgrade vm_page_prot
1574  * to the private version (using protection_map[] without the
1575  * VM_SHARED bit).
1576  */
1577 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1578 {
1579         vm_flags_t vm_flags = vma->vm_flags;
1580         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1581
1582         /* If it was private or non-writable, the write bit is already clear */
1583         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1584                 return 0;
1585
1586         /* The backer wishes to know when pages are first written to? */
1587         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1588                 return 1;
1589
1590         /* The open routine did something to the protections that pgprot_modify
1591          * won't preserve? */
1592         if (pgprot_val(vm_page_prot) !=
1593             pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1594                 return 0;
1595
1596         /* Do we need to track softdirty? */
1597         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1598                 return 1;
1599
1600         /* Specialty mapping? */
1601         if (vm_flags & VM_PFNMAP)
1602                 return 0;
1603
1604         /* Can the mapping track the dirty pages? */
1605         return vma->vm_file && vma->vm_file->f_mapping &&
1606                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1607 }
1608
1609 /*
1610  * We account for memory if it's a private writeable mapping,
1611  * not hugepages and VM_NORESERVE wasn't set.
1612  */
1613 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1614 {
1615         /*
1616          * hugetlb has its own accounting separate from the core VM
1617          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1618          */
1619         if (file && is_file_hugepages(file))
1620                 return 0;
1621
1622         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1623 }
1624
1625 unsigned long mmap_region(struct file *file, unsigned long addr,
1626                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1627                 struct list_head *uf)
1628 {
1629         struct mm_struct *mm = current->mm;
1630         struct vm_area_struct *vma, *prev;
1631         int error;
1632         struct rb_node **rb_link, *rb_parent;
1633         unsigned long charged = 0;
1634
1635         /* Check against address space limit. */
1636         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1637                 unsigned long nr_pages;
1638
1639                 /*
1640                  * MAP_FIXED may remove pages of mappings that intersects with
1641                  * requested mapping. Account for the pages it would unmap.
1642                  */
1643                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1644
1645                 if (!may_expand_vm(mm, vm_flags,
1646                                         (len >> PAGE_SHIFT) - nr_pages))
1647                         return -ENOMEM;
1648         }
1649
1650         /* Clear old maps */
1651         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1652                               &rb_parent)) {
1653                 if (do_munmap(mm, addr, len, uf))
1654                         return -ENOMEM;
1655         }
1656
1657         /*
1658          * Private writable mapping: check memory availability
1659          */
1660         if (accountable_mapping(file, vm_flags)) {
1661                 charged = len >> PAGE_SHIFT;
1662                 if (security_vm_enough_memory_mm(mm, charged))
1663                         return -ENOMEM;
1664                 vm_flags |= VM_ACCOUNT;
1665         }
1666
1667         /*
1668          * Can we just expand an old mapping?
1669          */
1670         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1671                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1672         if (vma)
1673                 goto out;
1674
1675         /*
1676          * Determine the object being mapped and call the appropriate
1677          * specific mapper. the address has already been validated, but
1678          * not unmapped, but the maps are removed from the list.
1679          */
1680         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1681         if (!vma) {
1682                 error = -ENOMEM;
1683                 goto unacct_error;
1684         }
1685
1686         vma->vm_mm = mm;
1687         vma->vm_start = addr;
1688         vma->vm_end = addr + len;
1689         vma->vm_flags = vm_flags;
1690         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1691         vma->vm_pgoff = pgoff;
1692         INIT_LIST_HEAD(&vma->anon_vma_chain);
1693
1694         if (file) {
1695                 if (vm_flags & VM_DENYWRITE) {
1696                         error = deny_write_access(file);
1697                         if (error)
1698                                 goto free_vma;
1699                 }
1700                 if (vm_flags & VM_SHARED) {
1701                         error = mapping_map_writable(file->f_mapping);
1702                         if (error)
1703                                 goto allow_write_and_free_vma;
1704                 }
1705
1706                 /* ->mmap() can change vma->vm_file, but must guarantee that
1707                  * vma_link() below can deny write-access if VM_DENYWRITE is set
1708                  * and map writably if VM_SHARED is set. This usually means the
1709                  * new file must not have been exposed to user-space, yet.
1710                  */
1711                 vma->vm_file = get_file(file);
1712                 error = call_mmap(file, vma);
1713                 if (error)
1714                         goto unmap_and_free_vma;
1715
1716                 /* Can addr have changed??
1717                  *
1718                  * Answer: Yes, several device drivers can do it in their
1719                  *         f_op->mmap method. -DaveM
1720                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1721                  *      be updated for vma_link()
1722                  */
1723                 WARN_ON_ONCE(addr != vma->vm_start);
1724
1725                 addr = vma->vm_start;
1726                 vm_flags = vma->vm_flags;
1727         } else if (vm_flags & VM_SHARED) {
1728                 error = shmem_zero_setup(vma);
1729                 if (error)
1730                         goto free_vma;
1731         }
1732
1733         vma_link(mm, vma, prev, rb_link, rb_parent);
1734         /* Once vma denies write, undo our temporary denial count */
1735         if (file) {
1736                 if (vm_flags & VM_SHARED)
1737                         mapping_unmap_writable(file->f_mapping);
1738                 if (vm_flags & VM_DENYWRITE)
1739                         allow_write_access(file);
1740         }
1741         file = vma->vm_file;
1742 out:
1743         perf_event_mmap(vma);
1744
1745         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1746         if (vm_flags & VM_LOCKED) {
1747                 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1748                                         vma == get_gate_vma(current->mm)))
1749                         mm->locked_vm += (len >> PAGE_SHIFT);
1750                 else
1751                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1752         }
1753
1754         if (file)
1755                 uprobe_mmap(vma);
1756
1757         /*
1758          * New (or expanded) vma always get soft dirty status.
1759          * Otherwise user-space soft-dirty page tracker won't
1760          * be able to distinguish situation when vma area unmapped,
1761          * then new mapped in-place (which must be aimed as
1762          * a completely new data area).
1763          */
1764         vma->vm_flags |= VM_SOFTDIRTY;
1765
1766         vma_set_page_prot(vma);
1767
1768         return addr;
1769
1770 unmap_and_free_vma:
1771         vma->vm_file = NULL;
1772         fput(file);
1773
1774         /* Undo any partial mapping done by a device driver. */
1775         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1776         charged = 0;
1777         if (vm_flags & VM_SHARED)
1778                 mapping_unmap_writable(file->f_mapping);
1779 allow_write_and_free_vma:
1780         if (vm_flags & VM_DENYWRITE)
1781                 allow_write_access(file);
1782 free_vma:
1783         kmem_cache_free(vm_area_cachep, vma);
1784 unacct_error:
1785         if (charged)
1786                 vm_unacct_memory(charged);
1787         return error;
1788 }
1789
1790 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1791 {
1792         /*
1793          * We implement the search by looking for an rbtree node that
1794          * immediately follows a suitable gap. That is,
1795          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1796          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1797          * - gap_end - gap_start >= length
1798          */
1799
1800         struct mm_struct *mm = current->mm;
1801         struct vm_area_struct *vma;
1802         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1803
1804         /* Adjust search length to account for worst case alignment overhead */
1805         length = info->length + info->align_mask;
1806         if (length < info->length)
1807                 return -ENOMEM;
1808
1809         /* Adjust search limits by the desired length */
1810         if (info->high_limit < length)
1811                 return -ENOMEM;
1812         high_limit = info->high_limit - length;
1813
1814         if (info->low_limit > high_limit)
1815                 return -ENOMEM;
1816         low_limit = info->low_limit + length;
1817
1818         /* Check if rbtree root looks promising */
1819         if (RB_EMPTY_ROOT(&mm->mm_rb))
1820                 goto check_highest;
1821         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1822         if (vma->rb_subtree_gap < length)
1823                 goto check_highest;
1824
1825         while (true) {
1826                 /* Visit left subtree if it looks promising */
1827                 gap_end = vm_start_gap(vma);
1828                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1829                         struct vm_area_struct *left =
1830                                 rb_entry(vma->vm_rb.rb_left,
1831                                          struct vm_area_struct, vm_rb);
1832                         if (left->rb_subtree_gap >= length) {
1833                                 vma = left;
1834                                 continue;
1835                         }
1836                 }
1837
1838                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1839 check_current:
1840                 /* Check if current node has a suitable gap */
1841                 if (gap_start > high_limit)
1842                         return -ENOMEM;
1843                 if (gap_end >= low_limit &&
1844                     gap_end > gap_start && gap_end - gap_start >= length)
1845                         goto found;
1846
1847                 /* Visit right subtree if it looks promising */
1848                 if (vma->vm_rb.rb_right) {
1849                         struct vm_area_struct *right =
1850                                 rb_entry(vma->vm_rb.rb_right,
1851                                          struct vm_area_struct, vm_rb);
1852                         if (right->rb_subtree_gap >= length) {
1853                                 vma = right;
1854                                 continue;
1855                         }
1856                 }
1857
1858                 /* Go back up the rbtree to find next candidate node */
1859                 while (true) {
1860                         struct rb_node *prev = &vma->vm_rb;
1861                         if (!rb_parent(prev))
1862                                 goto check_highest;
1863                         vma = rb_entry(rb_parent(prev),
1864                                        struct vm_area_struct, vm_rb);
1865                         if (prev == vma->vm_rb.rb_left) {
1866                                 gap_start = vm_end_gap(vma->vm_prev);
1867                                 gap_end = vm_start_gap(vma);
1868                                 goto check_current;
1869                         }
1870                 }
1871         }
1872
1873 check_highest:
1874         /* Check highest gap, which does not precede any rbtree node */
1875         gap_start = mm->highest_vm_end;
1876         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1877         if (gap_start > high_limit)
1878                 return -ENOMEM;
1879
1880 found:
1881         /* We found a suitable gap. Clip it with the original low_limit. */
1882         if (gap_start < info->low_limit)
1883                 gap_start = info->low_limit;
1884
1885         /* Adjust gap address to the desired alignment */
1886         gap_start += (info->align_offset - gap_start) & info->align_mask;
1887
1888         VM_BUG_ON(gap_start + info->length > info->high_limit);
1889         VM_BUG_ON(gap_start + info->length > gap_end);
1890         return gap_start;
1891 }
1892
1893 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1894 {
1895         struct mm_struct *mm = current->mm;
1896         struct vm_area_struct *vma;
1897         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1898
1899         /* Adjust search length to account for worst case alignment overhead */
1900         length = info->length + info->align_mask;
1901         if (length < info->length)
1902                 return -ENOMEM;
1903
1904         /*
1905          * Adjust search limits by the desired length.
1906          * See implementation comment at top of unmapped_area().
1907          */
1908         gap_end = info->high_limit;
1909         if (gap_end < length)
1910                 return -ENOMEM;
1911         high_limit = gap_end - length;
1912
1913         if (info->low_limit > high_limit)
1914                 return -ENOMEM;
1915         low_limit = info->low_limit + length;
1916
1917         /* Check highest gap, which does not precede any rbtree node */
1918         gap_start = mm->highest_vm_end;
1919         if (gap_start <= high_limit)
1920                 goto found_highest;
1921
1922         /* Check if rbtree root looks promising */
1923         if (RB_EMPTY_ROOT(&mm->mm_rb))
1924                 return -ENOMEM;
1925         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1926         if (vma->rb_subtree_gap < length)
1927                 return -ENOMEM;
1928
1929         while (true) {
1930                 /* Visit right subtree if it looks promising */
1931                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1932                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1933                         struct vm_area_struct *right =
1934                                 rb_entry(vma->vm_rb.rb_right,
1935                                          struct vm_area_struct, vm_rb);
1936                         if (right->rb_subtree_gap >= length) {
1937                                 vma = right;
1938                                 continue;
1939                         }
1940                 }
1941
1942 check_current:
1943                 /* Check if current node has a suitable gap */
1944                 gap_end = vm_start_gap(vma);
1945                 if (gap_end < low_limit)
1946                         return -ENOMEM;
1947                 if (gap_start <= high_limit &&
1948                     gap_end > gap_start && gap_end - gap_start >= length)
1949                         goto found;
1950
1951                 /* Visit left subtree if it looks promising */
1952                 if (vma->vm_rb.rb_left) {
1953                         struct vm_area_struct *left =
1954                                 rb_entry(vma->vm_rb.rb_left,
1955                                          struct vm_area_struct, vm_rb);
1956                         if (left->rb_subtree_gap >= length) {
1957                                 vma = left;
1958                                 continue;
1959                         }
1960                 }
1961
1962                 /* Go back up the rbtree to find next candidate node */
1963                 while (true) {
1964                         struct rb_node *prev = &vma->vm_rb;
1965                         if (!rb_parent(prev))
1966                                 return -ENOMEM;
1967                         vma = rb_entry(rb_parent(prev),
1968                                        struct vm_area_struct, vm_rb);
1969                         if (prev == vma->vm_rb.rb_right) {
1970                                 gap_start = vma->vm_prev ?
1971                                         vm_end_gap(vma->vm_prev) : 0;
1972                                 goto check_current;
1973                         }
1974                 }
1975         }
1976
1977 found:
1978         /* We found a suitable gap. Clip it with the original high_limit. */
1979         if (gap_end > info->high_limit)
1980                 gap_end = info->high_limit;
1981
1982 found_highest:
1983         /* Compute highest gap address at the desired alignment */
1984         gap_end -= info->length;
1985         gap_end -= (gap_end - info->align_offset) & info->align_mask;
1986
1987         VM_BUG_ON(gap_end < info->low_limit);
1988         VM_BUG_ON(gap_end < gap_start);
1989         return gap_end;
1990 }
1991
1992 /* Get an address range which is currently unmapped.
1993  * For shmat() with addr=0.
1994  *
1995  * Ugly calling convention alert:
1996  * Return value with the low bits set means error value,
1997  * ie
1998  *      if (ret & ~PAGE_MASK)
1999  *              error = ret;
2000  *
2001  * This function "knows" that -ENOMEM has the bits set.
2002  */
2003 #ifndef HAVE_ARCH_UNMAPPED_AREA
2004 unsigned long
2005 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2006                 unsigned long len, unsigned long pgoff, unsigned long flags)
2007 {
2008         struct mm_struct *mm = current->mm;
2009         struct vm_area_struct *vma, *prev;
2010         struct vm_unmapped_area_info info;
2011
2012         if (len > TASK_SIZE - mmap_min_addr)
2013                 return -ENOMEM;
2014
2015         if (flags & MAP_FIXED)
2016                 return addr;
2017
2018         if (addr) {
2019                 addr = PAGE_ALIGN(addr);
2020                 vma = find_vma_prev(mm, addr, &prev);
2021                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2022                     (!vma || addr + len <= vm_start_gap(vma)) &&
2023                     (!prev || addr >= vm_end_gap(prev)))
2024                         return addr;
2025         }
2026
2027         info.flags = 0;
2028         info.length = len;
2029         info.low_limit = mm->mmap_base;
2030         info.high_limit = TASK_SIZE;
2031         info.align_mask = 0;
2032         return vm_unmapped_area(&info);
2033 }
2034 #endif
2035
2036 /*
2037  * This mmap-allocator allocates new areas top-down from below the
2038  * stack's low limit (the base):
2039  */
2040 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2041 unsigned long
2042 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2043                           const unsigned long len, const unsigned long pgoff,
2044                           const unsigned long flags)
2045 {
2046         struct vm_area_struct *vma, *prev;
2047         struct mm_struct *mm = current->mm;
2048         unsigned long addr = addr0;
2049         struct vm_unmapped_area_info info;
2050
2051         /* requested length too big for entire address space */
2052         if (len > TASK_SIZE - mmap_min_addr)
2053                 return -ENOMEM;
2054
2055         if (flags & MAP_FIXED)
2056                 return addr;
2057
2058         /* requesting a specific address */
2059         if (addr) {
2060                 addr = PAGE_ALIGN(addr);
2061                 vma = find_vma_prev(mm, addr, &prev);
2062                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2063                                 (!vma || addr + len <= vm_start_gap(vma)) &&
2064                                 (!prev || addr >= vm_end_gap(prev)))
2065                         return addr;
2066         }
2067
2068         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2069         info.length = len;
2070         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2071         info.high_limit = mm->mmap_base;
2072         info.align_mask = 0;
2073         addr = vm_unmapped_area(&info);
2074
2075         /*
2076          * A failed mmap() very likely causes application failure,
2077          * so fall back to the bottom-up function here. This scenario
2078          * can happen with large stack limits and large mmap()
2079          * allocations.
2080          */
2081         if (offset_in_page(addr)) {
2082                 VM_BUG_ON(addr != -ENOMEM);
2083                 info.flags = 0;
2084                 info.low_limit = TASK_UNMAPPED_BASE;
2085                 info.high_limit = TASK_SIZE;
2086                 addr = vm_unmapped_area(&info);
2087         }
2088
2089         return addr;
2090 }
2091 #endif
2092
2093 unsigned long
2094 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2095                 unsigned long pgoff, unsigned long flags)
2096 {
2097         unsigned long (*get_area)(struct file *, unsigned long,
2098                                   unsigned long, unsigned long, unsigned long);
2099
2100         unsigned long error = arch_mmap_check(addr, len, flags);
2101         if (error)
2102                 return error;
2103
2104         /* Careful about overflows.. */
2105         if (len > TASK_SIZE)
2106                 return -ENOMEM;
2107
2108         get_area = current->mm->get_unmapped_area;
2109         if (file) {
2110                 if (file->f_op->get_unmapped_area)
2111                         get_area = file->f_op->get_unmapped_area;
2112         } else if (flags & MAP_SHARED) {
2113                 /*
2114                  * mmap_region() will call shmem_zero_setup() to create a file,
2115                  * so use shmem's get_unmapped_area in case it can be huge.
2116                  * do_mmap_pgoff() will clear pgoff, so match alignment.
2117                  */
2118                 pgoff = 0;
2119                 get_area = shmem_get_unmapped_area;
2120         }
2121
2122         addr = get_area(file, addr, len, pgoff, flags);
2123         if (IS_ERR_VALUE(addr))
2124                 return addr;
2125
2126         if (addr > TASK_SIZE - len)
2127                 return -ENOMEM;
2128         if (offset_in_page(addr))
2129                 return -EINVAL;
2130
2131         error = security_mmap_addr(addr);
2132         return error ? error : addr;
2133 }
2134
2135 EXPORT_SYMBOL(get_unmapped_area);
2136
2137 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2138 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2139 {
2140         struct rb_node *rb_node;
2141         struct vm_area_struct *vma;
2142
2143         /* Check the cache first. */
2144         vma = vmacache_find(mm, addr);
2145         if (likely(vma))
2146                 return vma;
2147
2148         rb_node = mm->mm_rb.rb_node;
2149
2150         while (rb_node) {
2151                 struct vm_area_struct *tmp;
2152
2153                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2154
2155                 if (tmp->vm_end > addr) {
2156                         vma = tmp;
2157                         if (tmp->vm_start <= addr)
2158                                 break;
2159                         rb_node = rb_node->rb_left;
2160                 } else
2161                         rb_node = rb_node->rb_right;
2162         }
2163
2164         if (vma)
2165                 vmacache_update(addr, vma);
2166         return vma;
2167 }
2168
2169 EXPORT_SYMBOL(find_vma);
2170
2171 /*
2172  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2173  */
2174 struct vm_area_struct *
2175 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2176                         struct vm_area_struct **pprev)
2177 {
2178         struct vm_area_struct *vma;
2179
2180         vma = find_vma(mm, addr);
2181         if (vma) {
2182                 *pprev = vma->vm_prev;
2183         } else {
2184                 struct rb_node *rb_node = mm->mm_rb.rb_node;
2185                 *pprev = NULL;
2186                 while (rb_node) {
2187                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2188                         rb_node = rb_node->rb_right;
2189                 }
2190         }
2191         return vma;
2192 }
2193
2194 /*
2195  * Verify that the stack growth is acceptable and
2196  * update accounting. This is shared with both the
2197  * grow-up and grow-down cases.
2198  */
2199 static int acct_stack_growth(struct vm_area_struct *vma,
2200                              unsigned long size, unsigned long grow)
2201 {
2202         struct mm_struct *mm = vma->vm_mm;
2203         unsigned long new_start;
2204
2205         /* address space limit tests */
2206         if (!may_expand_vm(mm, vma->vm_flags, grow))
2207                 return -ENOMEM;
2208
2209         /* Stack limit test */
2210         if (size > rlimit(RLIMIT_STACK))
2211                 return -ENOMEM;
2212
2213         /* mlock limit tests */
2214         if (vma->vm_flags & VM_LOCKED) {
2215                 unsigned long locked;
2216                 unsigned long limit;
2217                 locked = mm->locked_vm + grow;
2218                 limit = rlimit(RLIMIT_MEMLOCK);
2219                 limit >>= PAGE_SHIFT;
2220                 if (locked > limit && !capable(CAP_IPC_LOCK))
2221                         return -ENOMEM;
2222         }
2223
2224         /* Check to ensure the stack will not grow into a hugetlb-only region */
2225         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2226                         vma->vm_end - size;
2227         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2228                 return -EFAULT;
2229
2230         /*
2231          * Overcommit..  This must be the final test, as it will
2232          * update security statistics.
2233          */
2234         if (security_vm_enough_memory_mm(mm, grow))
2235                 return -ENOMEM;
2236
2237         return 0;
2238 }
2239
2240 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2241 /*
2242  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2243  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2244  */
2245 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2246 {
2247         struct mm_struct *mm = vma->vm_mm;
2248         struct vm_area_struct *next;
2249         unsigned long gap_addr;
2250         int error = 0;
2251
2252         if (!(vma->vm_flags & VM_GROWSUP))
2253                 return -EFAULT;
2254
2255         /* Guard against exceeding limits of the address space. */
2256         address &= PAGE_MASK;
2257         if (address >= (TASK_SIZE & PAGE_MASK))
2258                 return -ENOMEM;
2259         address += PAGE_SIZE;
2260
2261         /* Enforce stack_guard_gap */
2262         gap_addr = address + stack_guard_gap;
2263
2264         /* Guard against overflow */
2265         if (gap_addr < address || gap_addr > TASK_SIZE)
2266                 gap_addr = TASK_SIZE;
2267
2268         next = vma->vm_next;
2269         if (next && next->vm_start < gap_addr &&
2270                         (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2271                 if (!(next->vm_flags & VM_GROWSUP))
2272                         return -ENOMEM;
2273                 /* Check that both stack segments have the same anon_vma? */
2274         }
2275
2276         /* We must make sure the anon_vma is allocated. */
2277         if (unlikely(anon_vma_prepare(vma)))
2278                 return -ENOMEM;
2279
2280         /*
2281          * vma->vm_start/vm_end cannot change under us because the caller
2282          * is required to hold the mmap_sem in read mode.  We need the
2283          * anon_vma lock to serialize against concurrent expand_stacks.
2284          */
2285         anon_vma_lock_write(vma->anon_vma);
2286
2287         /* Somebody else might have raced and expanded it already */
2288         if (address > vma->vm_end) {
2289                 unsigned long size, grow;
2290
2291                 size = address - vma->vm_start;
2292                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2293
2294                 error = -ENOMEM;
2295                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2296                         error = acct_stack_growth(vma, size, grow);
2297                         if (!error) {
2298                                 /*
2299                                  * vma_gap_update() doesn't support concurrent
2300                                  * updates, but we only hold a shared mmap_sem
2301                                  * lock here, so we need to protect against
2302                                  * concurrent vma expansions.
2303                                  * anon_vma_lock_write() doesn't help here, as
2304                                  * we don't guarantee that all growable vmas
2305                                  * in a mm share the same root anon vma.
2306                                  * So, we reuse mm->page_table_lock to guard
2307                                  * against concurrent vma expansions.
2308                                  */
2309                                 spin_lock(&mm->page_table_lock);
2310                                 if (vma->vm_flags & VM_LOCKED)
2311                                         mm->locked_vm += grow;
2312                                 vm_stat_account(mm, vma->vm_flags, grow);
2313                                 anon_vma_interval_tree_pre_update_vma(vma);
2314                                 vma->vm_end = address;
2315                                 anon_vma_interval_tree_post_update_vma(vma);
2316                                 if (vma->vm_next)
2317                                         vma_gap_update(vma->vm_next);
2318                                 else
2319                                         mm->highest_vm_end = vm_end_gap(vma);
2320                                 spin_unlock(&mm->page_table_lock);
2321
2322                                 perf_event_mmap(vma);
2323                         }
2324                 }
2325         }
2326         anon_vma_unlock_write(vma->anon_vma);
2327         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2328         validate_mm(mm);
2329         return error;
2330 }
2331 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2332
2333 /*
2334  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2335  */
2336 int expand_downwards(struct vm_area_struct *vma,
2337                                    unsigned long address)
2338 {
2339         struct mm_struct *mm = vma->vm_mm;
2340         struct vm_area_struct *prev;
2341         int error;
2342
2343         address &= PAGE_MASK;
2344         error = security_mmap_addr(address);
2345         if (error)
2346                 return error;
2347
2348         /* Enforce stack_guard_gap */
2349         prev = vma->vm_prev;
2350         /* Check that both stack segments have the same anon_vma? */
2351         if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2352                         (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2353                 if (address - prev->vm_end < stack_guard_gap)
2354                         return -ENOMEM;
2355         }
2356
2357         /* We must make sure the anon_vma is allocated. */
2358         if (unlikely(anon_vma_prepare(vma)))
2359                 return -ENOMEM;
2360
2361         /*
2362          * vma->vm_start/vm_end cannot change under us because the caller
2363          * is required to hold the mmap_sem in read mode.  We need the
2364          * anon_vma lock to serialize against concurrent expand_stacks.
2365          */
2366         anon_vma_lock_write(vma->anon_vma);
2367
2368         /* Somebody else might have raced and expanded it already */
2369         if (address < vma->vm_start) {
2370                 unsigned long size, grow;
2371
2372                 size = vma->vm_end - address;
2373                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2374
2375                 error = -ENOMEM;
2376                 if (grow <= vma->vm_pgoff) {
2377                         error = acct_stack_growth(vma, size, grow);
2378                         if (!error) {
2379                                 /*
2380                                  * vma_gap_update() doesn't support concurrent
2381                                  * updates, but we only hold a shared mmap_sem
2382                                  * lock here, so we need to protect against
2383                                  * concurrent vma expansions.
2384                                  * anon_vma_lock_write() doesn't help here, as
2385                                  * we don't guarantee that all growable vmas
2386                                  * in a mm share the same root anon vma.
2387                                  * So, we reuse mm->page_table_lock to guard
2388                                  * against concurrent vma expansions.
2389                                  */
2390                                 spin_lock(&mm->page_table_lock);
2391                                 if (vma->vm_flags & VM_LOCKED)
2392                                         mm->locked_vm += grow;
2393                                 vm_stat_account(mm, vma->vm_flags, grow);
2394                                 anon_vma_interval_tree_pre_update_vma(vma);
2395                                 vma->vm_start = address;
2396                                 vma->vm_pgoff -= grow;
2397                                 anon_vma_interval_tree_post_update_vma(vma);
2398                                 vma_gap_update(vma);
2399                                 spin_unlock(&mm->page_table_lock);
2400
2401                                 perf_event_mmap(vma);
2402                         }
2403                 }
2404         }
2405         anon_vma_unlock_write(vma->anon_vma);
2406         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2407         validate_mm(mm);
2408         return error;
2409 }
2410
2411 /* enforced gap between the expanding stack and other mappings. */
2412 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2413
2414 static int __init cmdline_parse_stack_guard_gap(char *p)
2415 {
2416         unsigned long val;
2417         char *endptr;
2418
2419         val = simple_strtoul(p, &endptr, 10);
2420         if (!*endptr)
2421                 stack_guard_gap = val << PAGE_SHIFT;
2422
2423         return 0;
2424 }
2425 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2426
2427 #ifdef CONFIG_STACK_GROWSUP
2428 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2429 {
2430         return expand_upwards(vma, address);
2431 }
2432
2433 struct vm_area_struct *
2434 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2435 {
2436         struct vm_area_struct *vma, *prev;
2437
2438         addr &= PAGE_MASK;
2439         vma = find_vma_prev(mm, addr, &prev);
2440         if (vma && (vma->vm_start <= addr))
2441                 return vma;
2442         if (!prev || expand_stack(prev, addr))
2443                 return NULL;
2444         if (prev->vm_flags & VM_LOCKED)
2445                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2446         return prev;
2447 }
2448 #else
2449 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2450 {
2451         return expand_downwards(vma, address);
2452 }
2453
2454 struct vm_area_struct *
2455 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2456 {
2457         struct vm_area_struct *vma;
2458         unsigned long start;
2459
2460         addr &= PAGE_MASK;
2461         vma = find_vma(mm, addr);
2462         if (!vma)
2463                 return NULL;
2464         if (vma->vm_start <= addr)
2465                 return vma;
2466         if (!(vma->vm_flags & VM_GROWSDOWN))
2467                 return NULL;
2468         start = vma->vm_start;
2469         if (expand_stack(vma, addr))
2470                 return NULL;
2471         if (vma->vm_flags & VM_LOCKED)
2472                 populate_vma_page_range(vma, addr, start, NULL);
2473         return vma;
2474 }
2475 #endif
2476
2477 EXPORT_SYMBOL_GPL(find_extend_vma);
2478
2479 /*
2480  * Ok - we have the memory areas we should free on the vma list,
2481  * so release them, and do the vma updates.
2482  *
2483  * Called with the mm semaphore held.
2484  */
2485 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2486 {
2487         unsigned long nr_accounted = 0;
2488
2489         /* Update high watermark before we lower total_vm */
2490         update_hiwater_vm(mm);
2491         do {
2492                 long nrpages = vma_pages(vma);
2493
2494                 if (vma->vm_flags & VM_ACCOUNT)
2495                         nr_accounted += nrpages;
2496                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2497                 vma = remove_vma(vma);
2498         } while (vma);
2499         vm_unacct_memory(nr_accounted);
2500         validate_mm(mm);
2501 }
2502
2503 /*
2504  * Get rid of page table information in the indicated region.
2505  *
2506  * Called with the mm semaphore held.
2507  */
2508 static void unmap_region(struct mm_struct *mm,
2509                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2510                 unsigned long start, unsigned long end)
2511 {
2512         struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2513         struct mmu_gather tlb;
2514
2515         lru_add_drain();
2516         tlb_gather_mmu(&tlb, mm, start, end);
2517         update_hiwater_rss(mm);
2518         unmap_vmas(&tlb, vma, start, end);
2519         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2520                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2521         tlb_finish_mmu(&tlb, start, end);
2522 }
2523
2524 /*
2525  * Create a list of vma's touched by the unmap, removing them from the mm's
2526  * vma list as we go..
2527  */
2528 static void
2529 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2530         struct vm_area_struct *prev, unsigned long end)
2531 {
2532         struct vm_area_struct **insertion_point;
2533         struct vm_area_struct *tail_vma = NULL;
2534
2535         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2536         vma->vm_prev = NULL;
2537         do {
2538                 vma_rb_erase(vma, &mm->mm_rb);
2539                 mm->map_count--;
2540                 tail_vma = vma;
2541                 vma = vma->vm_next;
2542         } while (vma && vma->vm_start < end);
2543         *insertion_point = vma;
2544         if (vma) {
2545                 vma->vm_prev = prev;
2546                 vma_gap_update(vma);
2547         } else
2548                 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2549         tail_vma->vm_next = NULL;
2550
2551         /* Kill the cache */
2552         vmacache_invalidate(mm);
2553 }
2554
2555 /*
2556  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2557  * has already been checked or doesn't make sense to fail.
2558  */
2559 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2560                 unsigned long addr, int new_below)
2561 {
2562         struct vm_area_struct *new;
2563         int err;
2564
2565         if (vma->vm_ops && vma->vm_ops->split) {
2566                 err = vma->vm_ops->split(vma, addr);
2567                 if (err)
2568                         return err;
2569         }
2570
2571         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2572         if (!new)
2573                 return -ENOMEM;
2574
2575         /* most fields are the same, copy all, and then fixup */
2576         *new = *vma;
2577
2578         INIT_LIST_HEAD(&new->anon_vma_chain);
2579
2580         if (new_below)
2581                 new->vm_end = addr;
2582         else {
2583                 new->vm_start = addr;
2584                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2585         }
2586
2587         err = vma_dup_policy(vma, new);
2588         if (err)
2589                 goto out_free_vma;
2590
2591         err = anon_vma_clone(new, vma);
2592         if (err)
2593                 goto out_free_mpol;
2594
2595         if (new->vm_file)
2596                 get_file(new->vm_file);
2597
2598         if (new->vm_ops && new->vm_ops->open)
2599                 new->vm_ops->open(new);
2600
2601         if (new_below)
2602                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2603                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2604         else
2605                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2606
2607         /* Success. */
2608         if (!err)
2609                 return 0;
2610
2611         /* Clean everything up if vma_adjust failed. */
2612         if (new->vm_ops && new->vm_ops->close)
2613                 new->vm_ops->close(new);
2614         if (new->vm_file)
2615                 fput(new->vm_file);
2616         unlink_anon_vmas(new);
2617  out_free_mpol:
2618         mpol_put(vma_policy(new));
2619  out_free_vma:
2620         kmem_cache_free(vm_area_cachep, new);
2621         return err;
2622 }
2623
2624 /*
2625  * Split a vma into two pieces at address 'addr', a new vma is allocated
2626  * either for the first part or the tail.
2627  */
2628 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2629               unsigned long addr, int new_below)
2630 {
2631         if (mm->map_count >= sysctl_max_map_count)
2632                 return -ENOMEM;
2633
2634         return __split_vma(mm, vma, addr, new_below);
2635 }
2636
2637 /* Munmap is split into 2 main parts -- this part which finds
2638  * what needs doing, and the areas themselves, which do the
2639  * work.  This now handles partial unmappings.
2640  * Jeremy Fitzhardinge <jeremy@goop.org>
2641  */
2642 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2643               struct list_head *uf)
2644 {
2645         unsigned long end;
2646         struct vm_area_struct *vma, *prev, *last;
2647
2648         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2649                 return -EINVAL;
2650
2651         len = PAGE_ALIGN(len);
2652         if (len == 0)
2653                 return -EINVAL;
2654
2655         /* Find the first overlapping VMA */
2656         vma = find_vma(mm, start);
2657         if (!vma)
2658                 return 0;
2659         prev = vma->vm_prev;
2660         /* we have  start < vma->vm_end  */
2661
2662         /* if it doesn't overlap, we have nothing.. */
2663         end = start + len;
2664         if (vma->vm_start >= end)
2665                 return 0;
2666
2667         /*
2668          * If we need to split any vma, do it now to save pain later.
2669          *
2670          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2671          * unmapped vm_area_struct will remain in use: so lower split_vma
2672          * places tmp vma above, and higher split_vma places tmp vma below.
2673          */
2674         if (start > vma->vm_start) {
2675                 int error;
2676
2677                 /*
2678                  * Make sure that map_count on return from munmap() will
2679                  * not exceed its limit; but let map_count go just above
2680                  * its limit temporarily, to help free resources as expected.
2681                  */
2682                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2683                         return -ENOMEM;
2684
2685                 error = __split_vma(mm, vma, start, 0);
2686                 if (error)
2687                         return error;
2688                 prev = vma;
2689         }
2690
2691         /* Does it split the last one? */
2692         last = find_vma(mm, end);
2693         if (last && end > last->vm_start) {
2694                 int error = __split_vma(mm, last, end, 1);
2695                 if (error)
2696                         return error;
2697         }
2698         vma = prev ? prev->vm_next : mm->mmap;
2699
2700         if (unlikely(uf)) {
2701                 /*
2702                  * If userfaultfd_unmap_prep returns an error the vmas
2703                  * will remain splitted, but userland will get a
2704                  * highly unexpected error anyway. This is no
2705                  * different than the case where the first of the two
2706                  * __split_vma fails, but we don't undo the first
2707                  * split, despite we could. This is unlikely enough
2708                  * failure that it's not worth optimizing it for.
2709                  */
2710                 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2711                 if (error)
2712                         return error;
2713         }
2714
2715         /*
2716          * unlock any mlock()ed ranges before detaching vmas
2717          */
2718         if (mm->locked_vm) {
2719                 struct vm_area_struct *tmp = vma;
2720                 while (tmp && tmp->vm_start < end) {
2721                         if (tmp->vm_flags & VM_LOCKED) {
2722                                 mm->locked_vm -= vma_pages(tmp);
2723                                 munlock_vma_pages_all(tmp);
2724                         }
2725                         tmp = tmp->vm_next;
2726                 }
2727         }
2728
2729         /*
2730          * Remove the vma's, and unmap the actual pages
2731          */
2732         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2733         unmap_region(mm, vma, prev, start, end);
2734
2735         arch_unmap(mm, vma, start, end);
2736
2737         /* Fix up all other VM information */
2738         remove_vma_list(mm, vma);
2739
2740         return 0;
2741 }
2742
2743 int vm_munmap(unsigned long start, size_t len)
2744 {
2745         int ret;
2746         struct mm_struct *mm = current->mm;
2747         LIST_HEAD(uf);
2748
2749         if (down_write_killable(&mm->mmap_sem))
2750                 return -EINTR;
2751
2752         ret = do_munmap(mm, start, len, &uf);
2753         up_write(&mm->mmap_sem);
2754         userfaultfd_unmap_complete(mm, &uf);
2755         return ret;
2756 }
2757 EXPORT_SYMBOL(vm_munmap);
2758
2759 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2760 {
2761         profile_munmap(addr);
2762         return vm_munmap(addr, len);
2763 }
2764
2765
2766 /*
2767  * Emulation of deprecated remap_file_pages() syscall.
2768  */
2769 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2770                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2771 {
2772
2773         struct mm_struct *mm = current->mm;
2774         struct vm_area_struct *vma;
2775         unsigned long populate = 0;
2776         unsigned long ret = -EINVAL;
2777         struct file *file;
2778
2779         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2780                      current->comm, current->pid);
2781
2782         if (prot)
2783                 return ret;
2784         start = start & PAGE_MASK;
2785         size = size & PAGE_MASK;
2786
2787         if (start + size <= start)
2788                 return ret;
2789
2790         /* Does pgoff wrap? */
2791         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2792                 return ret;
2793
2794         if (down_write_killable(&mm->mmap_sem))
2795                 return -EINTR;
2796
2797         vma = find_vma(mm, start);
2798
2799         if (!vma || !(vma->vm_flags & VM_SHARED))
2800                 goto out;
2801
2802         if (start < vma->vm_start)
2803                 goto out;
2804
2805         if (start + size > vma->vm_end) {
2806                 struct vm_area_struct *next;
2807
2808                 for (next = vma->vm_next; next; next = next->vm_next) {
2809                         /* hole between vmas ? */
2810                         if (next->vm_start != next->vm_prev->vm_end)
2811                                 goto out;
2812
2813                         if (next->vm_file != vma->vm_file)
2814                                 goto out;
2815
2816                         if (next->vm_flags != vma->vm_flags)
2817                                 goto out;
2818
2819                         if (start + size <= next->vm_end)
2820                                 break;
2821                 }
2822
2823                 if (!next)
2824                         goto out;
2825         }
2826
2827         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2828         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2829         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2830
2831         flags &= MAP_NONBLOCK;
2832         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2833         if (vma->vm_flags & VM_LOCKED) {
2834                 struct vm_area_struct *tmp;
2835                 flags |= MAP_LOCKED;
2836
2837                 /* drop PG_Mlocked flag for over-mapped range */
2838                 for (tmp = vma; tmp->vm_start >= start + size;
2839                                 tmp = tmp->vm_next) {
2840                         /*
2841                          * Split pmd and munlock page on the border
2842                          * of the range.
2843                          */
2844                         vma_adjust_trans_huge(tmp, start, start + size, 0);
2845
2846                         munlock_vma_pages_range(tmp,
2847                                         max(tmp->vm_start, start),
2848                                         min(tmp->vm_end, start + size));
2849                 }
2850         }
2851
2852         file = get_file(vma->vm_file);
2853         ret = do_mmap_pgoff(vma->vm_file, start, size,
2854                         prot, flags, pgoff, &populate, NULL);
2855         fput(file);
2856 out:
2857         up_write(&mm->mmap_sem);
2858         if (populate)
2859                 mm_populate(ret, populate);
2860         if (!IS_ERR_VALUE(ret))
2861                 ret = 0;
2862         return ret;
2863 }
2864
2865 static inline void verify_mm_writelocked(struct mm_struct *mm)
2866 {
2867 #ifdef CONFIG_DEBUG_VM
2868         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2869                 WARN_ON(1);
2870                 up_read(&mm->mmap_sem);
2871         }
2872 #endif
2873 }
2874
2875 /*
2876  *  this is really a simplified "do_mmap".  it only handles
2877  *  anonymous maps.  eventually we may be able to do some
2878  *  brk-specific accounting here.
2879  */
2880 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags, struct list_head *uf)
2881 {
2882         struct mm_struct *mm = current->mm;
2883         struct vm_area_struct *vma, *prev;
2884         unsigned long len;
2885         struct rb_node **rb_link, *rb_parent;
2886         pgoff_t pgoff = addr >> PAGE_SHIFT;
2887         int error;
2888
2889         len = PAGE_ALIGN(request);
2890         if (len < request)
2891                 return -ENOMEM;
2892         if (!len)
2893                 return 0;
2894
2895         /* Until we need other flags, refuse anything except VM_EXEC. */
2896         if ((flags & (~VM_EXEC)) != 0)
2897                 return -EINVAL;
2898         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2899
2900         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2901         if (offset_in_page(error))
2902                 return error;
2903
2904         error = mlock_future_check(mm, mm->def_flags, len);
2905         if (error)
2906                 return error;
2907
2908         /*
2909          * mm->mmap_sem is required to protect against another thread
2910          * changing the mappings in case we sleep.
2911          */
2912         verify_mm_writelocked(mm);
2913
2914         /*
2915          * Clear old maps.  this also does some error checking for us
2916          */
2917         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2918                               &rb_parent)) {
2919                 if (do_munmap(mm, addr, len, uf))
2920                         return -ENOMEM;
2921         }
2922
2923         /* Check against address space limits *after* clearing old maps... */
2924         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2925                 return -ENOMEM;
2926
2927         if (mm->map_count > sysctl_max_map_count)
2928                 return -ENOMEM;
2929
2930         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2931                 return -ENOMEM;
2932
2933         /* Can we just expand an old private anonymous mapping? */
2934         vma = vma_merge(mm, prev, addr, addr + len, flags,
2935                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2936         if (vma)
2937                 goto out;
2938
2939         /*
2940          * create a vma struct for an anonymous mapping
2941          */
2942         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2943         if (!vma) {
2944                 vm_unacct_memory(len >> PAGE_SHIFT);
2945                 return -ENOMEM;
2946         }
2947
2948         INIT_LIST_HEAD(&vma->anon_vma_chain);
2949         vma->vm_mm = mm;
2950         vma->vm_start = addr;
2951         vma->vm_end = addr + len;
2952         vma->vm_pgoff = pgoff;
2953         vma->vm_flags = flags;
2954         vma->vm_page_prot = vm_get_page_prot(flags);
2955         vma_link(mm, vma, prev, rb_link, rb_parent);
2956 out:
2957         perf_event_mmap(vma);
2958         mm->total_vm += len >> PAGE_SHIFT;
2959         mm->data_vm += len >> PAGE_SHIFT;
2960         if (flags & VM_LOCKED)
2961                 mm->locked_vm += (len >> PAGE_SHIFT);
2962         vma->vm_flags |= VM_SOFTDIRTY;
2963         return 0;
2964 }
2965
2966 static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf)
2967 {
2968         return do_brk_flags(addr, len, 0, uf);
2969 }
2970
2971 int vm_brk_flags(unsigned long addr, unsigned long len, unsigned long flags)
2972 {
2973         struct mm_struct *mm = current->mm;
2974         int ret;
2975         bool populate;
2976         LIST_HEAD(uf);
2977
2978         if (down_write_killable(&mm->mmap_sem))
2979                 return -EINTR;
2980
2981         ret = do_brk_flags(addr, len, flags, &uf);
2982         populate = ((mm->def_flags & VM_LOCKED) != 0);
2983         up_write(&mm->mmap_sem);
2984         userfaultfd_unmap_complete(mm, &uf);
2985         if (populate && !ret)
2986                 mm_populate(addr, len);
2987         return ret;
2988 }
2989 EXPORT_SYMBOL(vm_brk_flags);
2990
2991 int vm_brk(unsigned long addr, unsigned long len)
2992 {
2993         return vm_brk_flags(addr, len, 0);
2994 }
2995 EXPORT_SYMBOL(vm_brk);
2996
2997 /* Release all mmaps. */
2998 void exit_mmap(struct mm_struct *mm)
2999 {
3000         struct mmu_gather tlb;
3001         struct vm_area_struct *vma;
3002         unsigned long nr_accounted = 0;
3003
3004         /* mm's last user has gone, and its about to be pulled down */
3005         mmu_notifier_release(mm);
3006
3007         if (mm->locked_vm) {
3008                 vma = mm->mmap;
3009                 while (vma) {
3010                         if (vma->vm_flags & VM_LOCKED)
3011                                 munlock_vma_pages_all(vma);
3012                         vma = vma->vm_next;
3013                 }
3014         }
3015
3016         arch_exit_mmap(mm);
3017
3018         vma = mm->mmap;
3019         if (!vma)       /* Can happen if dup_mmap() received an OOM */
3020                 return;
3021
3022         lru_add_drain();
3023         flush_cache_mm(mm);
3024         tlb_gather_mmu(&tlb, mm, 0, -1);
3025         /* update_hiwater_rss(mm) here? but nobody should be looking */
3026         /* Use -1 here to ensure all VMAs in the mm are unmapped */
3027         unmap_vmas(&tlb, vma, 0, -1);
3028
3029         if (unlikely(mm_is_oom_victim(mm))) {
3030                 /*
3031                  * Wait for oom_reap_task() to stop working on this
3032                  * mm. Because MMF_OOM_SKIP is already set before
3033                  * calling down_read(), oom_reap_task() will not run
3034                  * on this "mm" post up_write().
3035                  *
3036                  * mm_is_oom_victim() cannot be set from under us
3037                  * either because victim->mm is already set to NULL
3038                  * under task_lock before calling mmput and oom_mm is
3039                  * set not NULL by the OOM killer only if victim->mm
3040                  * is found not NULL while holding the task_lock.
3041                  */
3042                 set_bit(MMF_OOM_SKIP, &mm->flags);
3043                 down_write(&mm->mmap_sem);
3044                 up_write(&mm->mmap_sem);
3045         }
3046         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3047         tlb_finish_mmu(&tlb, 0, -1);
3048
3049         /*
3050          * Walk the list again, actually closing and freeing it,
3051          * with preemption enabled, without holding any MM locks.
3052          */
3053         while (vma) {
3054                 if (vma->vm_flags & VM_ACCOUNT)
3055                         nr_accounted += vma_pages(vma);
3056                 vma = remove_vma(vma);
3057         }
3058         vm_unacct_memory(nr_accounted);
3059 }
3060
3061 /* Insert vm structure into process list sorted by address
3062  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3063  * then i_mmap_rwsem is taken here.
3064  */
3065 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3066 {
3067         struct vm_area_struct *prev;
3068         struct rb_node **rb_link, *rb_parent;
3069
3070         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3071                            &prev, &rb_link, &rb_parent))
3072                 return -ENOMEM;
3073         if ((vma->vm_flags & VM_ACCOUNT) &&
3074              security_vm_enough_memory_mm(mm, vma_pages(vma)))
3075                 return -ENOMEM;
3076
3077         /*
3078          * The vm_pgoff of a purely anonymous vma should be irrelevant
3079          * until its first write fault, when page's anon_vma and index
3080          * are set.  But now set the vm_pgoff it will almost certainly
3081          * end up with (unless mremap moves it elsewhere before that
3082          * first wfault), so /proc/pid/maps tells a consistent story.
3083          *
3084          * By setting it to reflect the virtual start address of the
3085          * vma, merges and splits can happen in a seamless way, just
3086          * using the existing file pgoff checks and manipulations.
3087          * Similarly in do_mmap_pgoff and in do_brk.
3088          */
3089         if (vma_is_anonymous(vma)) {
3090                 BUG_ON(vma->anon_vma);
3091                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3092         }
3093
3094         vma_link(mm, vma, prev, rb_link, rb_parent);
3095         return 0;
3096 }
3097
3098 /*
3099  * Copy the vma structure to a new location in the same mm,
3100  * prior to moving page table entries, to effect an mremap move.
3101  */
3102 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3103         unsigned long addr, unsigned long len, pgoff_t pgoff,
3104         bool *need_rmap_locks)
3105 {
3106         struct vm_area_struct *vma = *vmap;
3107         unsigned long vma_start = vma->vm_start;
3108         struct mm_struct *mm = vma->vm_mm;
3109         struct vm_area_struct *new_vma, *prev;
3110         struct rb_node **rb_link, *rb_parent;
3111         bool faulted_in_anon_vma = true;
3112
3113         /*
3114          * If anonymous vma has not yet been faulted, update new pgoff
3115          * to match new location, to increase its chance of merging.
3116          */
3117         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3118                 pgoff = addr >> PAGE_SHIFT;
3119                 faulted_in_anon_vma = false;
3120         }
3121
3122         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3123                 return NULL;    /* should never get here */
3124         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3125                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3126                             vma->vm_userfaultfd_ctx);
3127         if (new_vma) {
3128                 /*
3129                  * Source vma may have been merged into new_vma
3130                  */
3131                 if (unlikely(vma_start >= new_vma->vm_start &&
3132                              vma_start < new_vma->vm_end)) {
3133                         /*
3134                          * The only way we can get a vma_merge with
3135                          * self during an mremap is if the vma hasn't
3136                          * been faulted in yet and we were allowed to
3137                          * reset the dst vma->vm_pgoff to the
3138                          * destination address of the mremap to allow
3139                          * the merge to happen. mremap must change the
3140                          * vm_pgoff linearity between src and dst vmas
3141                          * (in turn preventing a vma_merge) to be
3142                          * safe. It is only safe to keep the vm_pgoff
3143                          * linear if there are no pages mapped yet.
3144                          */
3145                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3146                         *vmap = vma = new_vma;
3147                 }
3148                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3149         } else {
3150                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3151                 if (!new_vma)
3152                         goto out;
3153                 *new_vma = *vma;
3154                 new_vma->vm_start = addr;
3155                 new_vma->vm_end = addr + len;
3156                 new_vma->vm_pgoff = pgoff;
3157                 if (vma_dup_policy(vma, new_vma))
3158                         goto out_free_vma;
3159                 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3160                 if (anon_vma_clone(new_vma, vma))
3161                         goto out_free_mempol;
3162                 if (new_vma->vm_file)
3163                         get_file(new_vma->vm_file);
3164                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3165                         new_vma->vm_ops->open(new_vma);
3166                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3167                 *need_rmap_locks = false;
3168         }
3169         return new_vma;
3170
3171 out_free_mempol:
3172         mpol_put(vma_policy(new_vma));
3173 out_free_vma:
3174         kmem_cache_free(vm_area_cachep, new_vma);
3175 out:
3176         return NULL;
3177 }
3178
3179 /*
3180  * Return true if the calling process may expand its vm space by the passed
3181  * number of pages
3182  */
3183 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3184 {
3185         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3186                 return false;
3187
3188         if (is_data_mapping(flags) &&
3189             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3190                 /* Workaround for Valgrind */
3191                 if (rlimit(RLIMIT_DATA) == 0 &&
3192                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3193                         return true;
3194
3195                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3196                              current->comm, current->pid,
3197                              (mm->data_vm + npages) << PAGE_SHIFT,
3198                              rlimit(RLIMIT_DATA),
3199                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3200
3201                 if (!ignore_rlimit_data)
3202                         return false;
3203         }
3204
3205         return true;
3206 }
3207
3208 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3209 {
3210         mm->total_vm += npages;
3211
3212         if (is_exec_mapping(flags))
3213                 mm->exec_vm += npages;
3214         else if (is_stack_mapping(flags))
3215                 mm->stack_vm += npages;
3216         else if (is_data_mapping(flags))
3217                 mm->data_vm += npages;
3218 }
3219
3220 static int special_mapping_fault(struct vm_fault *vmf);
3221
3222 /*
3223  * Having a close hook prevents vma merging regardless of flags.
3224  */
3225 static void special_mapping_close(struct vm_area_struct *vma)
3226 {
3227 }
3228
3229 static const char *special_mapping_name(struct vm_area_struct *vma)
3230 {
3231         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3232 }
3233
3234 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3235 {
3236         struct vm_special_mapping *sm = new_vma->vm_private_data;
3237
3238         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3239                 return -EFAULT;
3240
3241         if (sm->mremap)
3242                 return sm->mremap(sm, new_vma);
3243
3244         return 0;
3245 }
3246
3247 static const struct vm_operations_struct special_mapping_vmops = {
3248         .close = special_mapping_close,
3249         .fault = special_mapping_fault,
3250         .mremap = special_mapping_mremap,
3251         .name = special_mapping_name,
3252 };
3253
3254 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3255         .close = special_mapping_close,
3256         .fault = special_mapping_fault,
3257 };
3258
3259 static int special_mapping_fault(struct vm_fault *vmf)
3260 {
3261         struct vm_area_struct *vma = vmf->vma;
3262         pgoff_t pgoff;
3263         struct page **pages;
3264
3265         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3266                 pages = vma->vm_private_data;
3267         } else {
3268                 struct vm_special_mapping *sm = vma->vm_private_data;
3269
3270                 if (sm->fault)
3271                         return sm->fault(sm, vmf->vma, vmf);
3272
3273                 pages = sm->pages;
3274         }
3275
3276         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3277                 pgoff--;
3278
3279         if (*pages) {
3280                 struct page *page = *pages;
3281                 get_page(page);
3282                 vmf->page = page;
3283                 return 0;
3284         }
3285
3286         return VM_FAULT_SIGBUS;
3287 }
3288
3289 static struct vm_area_struct *__install_special_mapping(
3290         struct mm_struct *mm,
3291         unsigned long addr, unsigned long len,
3292         unsigned long vm_flags, void *priv,
3293         const struct vm_operations_struct *ops)
3294 {
3295         int ret;
3296         struct vm_area_struct *vma;
3297
3298         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3299         if (unlikely(vma == NULL))
3300                 return ERR_PTR(-ENOMEM);
3301
3302         INIT_LIST_HEAD(&vma->anon_vma_chain);
3303         vma->vm_mm = mm;
3304         vma->vm_start = addr;
3305         vma->vm_end = addr + len;
3306
3307         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3308         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3309
3310         vma->vm_ops = ops;
3311         vma->vm_private_data = priv;
3312
3313         ret = insert_vm_struct(mm, vma);
3314         if (ret)
3315                 goto out;
3316
3317         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3318
3319         perf_event_mmap(vma);
3320
3321         return vma;
3322
3323 out:
3324         kmem_cache_free(vm_area_cachep, vma);
3325         return ERR_PTR(ret);
3326 }
3327
3328 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3329         const struct vm_special_mapping *sm)
3330 {
3331         return vma->vm_private_data == sm &&
3332                 (vma->vm_ops == &special_mapping_vmops ||
3333                  vma->vm_ops == &legacy_special_mapping_vmops);
3334 }
3335
3336 /*
3337  * Called with mm->mmap_sem held for writing.
3338  * Insert a new vma covering the given region, with the given flags.
3339  * Its pages are supplied by the given array of struct page *.
3340  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3341  * The region past the last page supplied will always produce SIGBUS.
3342  * The array pointer and the pages it points to are assumed to stay alive
3343  * for as long as this mapping might exist.
3344  */
3345 struct vm_area_struct *_install_special_mapping(
3346         struct mm_struct *mm,
3347         unsigned long addr, unsigned long len,
3348         unsigned long vm_flags, const struct vm_special_mapping *spec)
3349 {
3350         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3351                                         &special_mapping_vmops);
3352 }
3353
3354 int install_special_mapping(struct mm_struct *mm,
3355                             unsigned long addr, unsigned long len,
3356                             unsigned long vm_flags, struct page **pages)
3357 {
3358         struct vm_area_struct *vma = __install_special_mapping(
3359                 mm, addr, len, vm_flags, (void *)pages,
3360                 &legacy_special_mapping_vmops);
3361
3362         return PTR_ERR_OR_ZERO(vma);
3363 }
3364
3365 static DEFINE_MUTEX(mm_all_locks_mutex);
3366
3367 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3368 {
3369         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3370                 /*
3371                  * The LSB of head.next can't change from under us
3372                  * because we hold the mm_all_locks_mutex.
3373                  */
3374                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3375                 /*
3376                  * We can safely modify head.next after taking the
3377                  * anon_vma->root->rwsem. If some other vma in this mm shares
3378                  * the same anon_vma we won't take it again.
3379                  *
3380                  * No need of atomic instructions here, head.next
3381                  * can't change from under us thanks to the
3382                  * anon_vma->root->rwsem.
3383                  */
3384                 if (__test_and_set_bit(0, (unsigned long *)
3385                                        &anon_vma->root->rb_root.rb_root.rb_node))
3386                         BUG();
3387         }
3388 }
3389
3390 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3391 {
3392         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3393                 /*
3394                  * AS_MM_ALL_LOCKS can't change from under us because
3395                  * we hold the mm_all_locks_mutex.
3396                  *
3397                  * Operations on ->flags have to be atomic because
3398                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3399                  * mm_all_locks_mutex, there may be other cpus
3400                  * changing other bitflags in parallel to us.
3401                  */
3402                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3403                         BUG();
3404                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3405         }
3406 }
3407
3408 /*
3409  * This operation locks against the VM for all pte/vma/mm related
3410  * operations that could ever happen on a certain mm. This includes
3411  * vmtruncate, try_to_unmap, and all page faults.
3412  *
3413  * The caller must take the mmap_sem in write mode before calling
3414  * mm_take_all_locks(). The caller isn't allowed to release the
3415  * mmap_sem until mm_drop_all_locks() returns.
3416  *
3417  * mmap_sem in write mode is required in order to block all operations
3418  * that could modify pagetables and free pages without need of
3419  * altering the vma layout. It's also needed in write mode to avoid new
3420  * anon_vmas to be associated with existing vmas.
3421  *
3422  * A single task can't take more than one mm_take_all_locks() in a row
3423  * or it would deadlock.
3424  *
3425  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3426  * mapping->flags avoid to take the same lock twice, if more than one
3427  * vma in this mm is backed by the same anon_vma or address_space.
3428  *
3429  * We take locks in following order, accordingly to comment at beginning
3430  * of mm/rmap.c:
3431  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3432  *     hugetlb mapping);
3433  *   - all i_mmap_rwsem locks;
3434  *   - all anon_vma->rwseml
3435  *
3436  * We can take all locks within these types randomly because the VM code
3437  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3438  * mm_all_locks_mutex.
3439  *
3440  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3441  * that may have to take thousand of locks.
3442  *
3443  * mm_take_all_locks() can fail if it's interrupted by signals.
3444  */
3445 int mm_take_all_locks(struct mm_struct *mm)
3446 {
3447         struct vm_area_struct *vma;
3448         struct anon_vma_chain *avc;
3449
3450         BUG_ON(down_read_trylock(&mm->mmap_sem));
3451
3452         mutex_lock(&mm_all_locks_mutex);
3453
3454         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3455                 if (signal_pending(current))
3456                         goto out_unlock;
3457                 if (vma->vm_file && vma->vm_file->f_mapping &&
3458                                 is_vm_hugetlb_page(vma))
3459                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3460         }
3461
3462         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3463                 if (signal_pending(current))
3464                         goto out_unlock;