mm: introduce MAP_FIXED_NOREPLACE
[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         /* force arch specific MAP_FIXED handling in get_unmapped_area */
1346         if (flags & MAP_FIXED_NOREPLACE)
1347                 flags |= MAP_FIXED;
1348
1349         if (!(flags & MAP_FIXED))
1350                 addr = round_hint_to_min(addr);
1351
1352         /* Careful about overflows.. */
1353         len = PAGE_ALIGN(len);
1354         if (!len)
1355                 return -ENOMEM;
1356
1357         /* offset overflow? */
1358         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1359                 return -EOVERFLOW;
1360
1361         /* Too many mappings? */
1362         if (mm->map_count > sysctl_max_map_count)
1363                 return -ENOMEM;
1364
1365         /* Obtain the address to map to. we verify (or select) it and ensure
1366          * that it represents a valid section of the address space.
1367          */
1368         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1369         if (offset_in_page(addr))
1370                 return addr;
1371
1372         if (flags & MAP_FIXED_NOREPLACE) {
1373                 struct vm_area_struct *vma = find_vma(mm, addr);
1374
1375                 if (vma && vma->vm_start <= addr)
1376                         return -EEXIST;
1377         }
1378
1379         if (prot == PROT_EXEC) {
1380                 pkey = execute_only_pkey(mm);
1381                 if (pkey < 0)
1382                         pkey = 0;
1383         }
1384
1385         /* Do simple checking here so the lower-level routines won't have
1386          * to. we assume access permissions have been handled by the open
1387          * of the memory object, so we don't do any here.
1388          */
1389         vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1390                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1391
1392         if (flags & MAP_LOCKED)
1393                 if (!can_do_mlock())
1394                         return -EPERM;
1395
1396         if (mlock_future_check(mm, vm_flags, len))
1397                 return -EAGAIN;
1398
1399         if (file) {
1400                 struct inode *inode = file_inode(file);
1401                 unsigned long flags_mask;
1402
1403                 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1404
1405                 switch (flags & MAP_TYPE) {
1406                 case MAP_SHARED:
1407                         /*
1408                          * Force use of MAP_SHARED_VALIDATE with non-legacy
1409                          * flags. E.g. MAP_SYNC is dangerous to use with
1410                          * MAP_SHARED as you don't know which consistency model
1411                          * you will get. We silently ignore unsupported flags
1412                          * with MAP_SHARED to preserve backward compatibility.
1413                          */
1414                         flags &= LEGACY_MAP_MASK;
1415                         /* fall through */
1416                 case MAP_SHARED_VALIDATE:
1417                         if (flags & ~flags_mask)
1418                                 return -EOPNOTSUPP;
1419                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1420                                 return -EACCES;
1421
1422                         /*
1423                          * Make sure we don't allow writing to an append-only
1424                          * file..
1425                          */
1426                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1427                                 return -EACCES;
1428
1429                         /*
1430                          * Make sure there are no mandatory locks on the file.
1431                          */
1432                         if (locks_verify_locked(file))
1433                                 return -EAGAIN;
1434
1435                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1436                         if (!(file->f_mode & FMODE_WRITE))
1437                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1438
1439                         /* fall through */
1440                 case MAP_PRIVATE:
1441                         if (!(file->f_mode & FMODE_READ))
1442                                 return -EACCES;
1443                         if (path_noexec(&file->f_path)) {
1444                                 if (vm_flags & VM_EXEC)
1445                                         return -EPERM;
1446                                 vm_flags &= ~VM_MAYEXEC;
1447                         }
1448
1449                         if (!file->f_op->mmap)
1450                                 return -ENODEV;
1451                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1452                                 return -EINVAL;
1453                         break;
1454
1455                 default:
1456                         return -EINVAL;
1457                 }
1458         } else {
1459                 switch (flags & MAP_TYPE) {
1460                 case MAP_SHARED:
1461                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1462                                 return -EINVAL;
1463                         /*
1464                          * Ignore pgoff.
1465                          */
1466                         pgoff = 0;
1467                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1468                         break;
1469                 case MAP_PRIVATE:
1470                         /*
1471                          * Set pgoff according to addr for anon_vma.
1472                          */
1473                         pgoff = addr >> PAGE_SHIFT;
1474                         break;
1475                 default:
1476                         return -EINVAL;
1477                 }
1478         }
1479
1480         /*
1481          * Set 'VM_NORESERVE' if we should not account for the
1482          * memory use of this mapping.
1483          */
1484         if (flags & MAP_NORESERVE) {
1485                 /* We honor MAP_NORESERVE if allowed to overcommit */
1486                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1487                         vm_flags |= VM_NORESERVE;
1488
1489                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1490                 if (file && is_file_hugepages(file))
1491                         vm_flags |= VM_NORESERVE;
1492         }
1493
1494         addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1495         if (!IS_ERR_VALUE(addr) &&
1496             ((vm_flags & VM_LOCKED) ||
1497              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1498                 *populate = len;
1499         return addr;
1500 }
1501
1502 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1503                               unsigned long prot, unsigned long flags,
1504                               unsigned long fd, unsigned long pgoff)
1505 {
1506         struct file *file = NULL;
1507         unsigned long retval;
1508
1509         if (!(flags & MAP_ANONYMOUS)) {
1510                 audit_mmap_fd(fd, flags);
1511                 file = fget(fd);
1512                 if (!file)
1513                         return -EBADF;
1514                 if (is_file_hugepages(file))
1515                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1516                 retval = -EINVAL;
1517                 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1518                         goto out_fput;
1519         } else if (flags & MAP_HUGETLB) {
1520                 struct user_struct *user = NULL;
1521                 struct hstate *hs;
1522
1523                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1524                 if (!hs)
1525                         return -EINVAL;
1526
1527                 len = ALIGN(len, huge_page_size(hs));
1528                 /*
1529                  * VM_NORESERVE is used because the reservations will be
1530                  * taken when vm_ops->mmap() is called
1531                  * A dummy user value is used because we are not locking
1532                  * memory so no accounting is necessary
1533                  */
1534                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1535                                 VM_NORESERVE,
1536                                 &user, HUGETLB_ANONHUGE_INODE,
1537                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1538                 if (IS_ERR(file))
1539                         return PTR_ERR(file);
1540         }
1541
1542         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1543
1544         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1545 out_fput:
1546         if (file)
1547                 fput(file);
1548         return retval;
1549 }
1550
1551 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1552                 unsigned long, prot, unsigned long, flags,
1553                 unsigned long, fd, unsigned long, pgoff)
1554 {
1555         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1556 }
1557
1558 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1559 struct mmap_arg_struct {
1560         unsigned long addr;
1561         unsigned long len;
1562         unsigned long prot;
1563         unsigned long flags;
1564         unsigned long fd;
1565         unsigned long offset;
1566 };
1567
1568 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1569 {
1570         struct mmap_arg_struct a;
1571
1572         if (copy_from_user(&a, arg, sizeof(a)))
1573                 return -EFAULT;
1574         if (offset_in_page(a.offset))
1575                 return -EINVAL;
1576
1577         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1578                                a.offset >> PAGE_SHIFT);
1579 }
1580 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1581
1582 /*
1583  * Some shared mappigns will want the pages marked read-only
1584  * to track write events. If so, we'll downgrade vm_page_prot
1585  * to the private version (using protection_map[] without the
1586  * VM_SHARED bit).
1587  */
1588 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1589 {
1590         vm_flags_t vm_flags = vma->vm_flags;
1591         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1592
1593         /* If it was private or non-writable, the write bit is already clear */
1594         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1595                 return 0;
1596
1597         /* The backer wishes to know when pages are first written to? */
1598         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1599                 return 1;
1600
1601         /* The open routine did something to the protections that pgprot_modify
1602          * won't preserve? */
1603         if (pgprot_val(vm_page_prot) !=
1604             pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1605                 return 0;
1606
1607         /* Do we need to track softdirty? */
1608         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1609                 return 1;
1610
1611         /* Specialty mapping? */
1612         if (vm_flags & VM_PFNMAP)
1613                 return 0;
1614
1615         /* Can the mapping track the dirty pages? */
1616         return vma->vm_file && vma->vm_file->f_mapping &&
1617                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1618 }
1619
1620 /*
1621  * We account for memory if it's a private writeable mapping,
1622  * not hugepages and VM_NORESERVE wasn't set.
1623  */
1624 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1625 {
1626         /*
1627          * hugetlb has its own accounting separate from the core VM
1628          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1629          */
1630         if (file && is_file_hugepages(file))
1631                 return 0;
1632
1633         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1634 }
1635
1636 unsigned long mmap_region(struct file *file, unsigned long addr,
1637                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1638                 struct list_head *uf)
1639 {
1640         struct mm_struct *mm = current->mm;
1641         struct vm_area_struct *vma, *prev;
1642         int error;
1643         struct rb_node **rb_link, *rb_parent;
1644         unsigned long charged = 0;
1645
1646         /* Check against address space limit. */
1647         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1648                 unsigned long nr_pages;
1649
1650                 /*
1651                  * MAP_FIXED may remove pages of mappings that intersects with
1652                  * requested mapping. Account for the pages it would unmap.
1653                  */
1654                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1655
1656                 if (!may_expand_vm(mm, vm_flags,
1657                                         (len >> PAGE_SHIFT) - nr_pages))
1658                         return -ENOMEM;
1659         }
1660
1661         /* Clear old maps */
1662         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1663                               &rb_parent)) {
1664                 if (do_munmap(mm, addr, len, uf))
1665                         return -ENOMEM;
1666         }
1667
1668         /*
1669          * Private writable mapping: check memory availability
1670          */
1671         if (accountable_mapping(file, vm_flags)) {
1672                 charged = len >> PAGE_SHIFT;
1673                 if (security_vm_enough_memory_mm(mm, charged))
1674                         return -ENOMEM;
1675                 vm_flags |= VM_ACCOUNT;
1676         }
1677
1678         /*
1679          * Can we just expand an old mapping?
1680          */
1681         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1682                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1683         if (vma)
1684                 goto out;
1685
1686         /*
1687          * Determine the object being mapped and call the appropriate
1688          * specific mapper. the address has already been validated, but
1689          * not unmapped, but the maps are removed from the list.
1690          */
1691         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1692         if (!vma) {
1693                 error = -ENOMEM;
1694                 goto unacct_error;
1695         }
1696
1697         vma->vm_mm = mm;
1698         vma->vm_start = addr;
1699         vma->vm_end = addr + len;
1700         vma->vm_flags = vm_flags;
1701         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1702         vma->vm_pgoff = pgoff;
1703         INIT_LIST_HEAD(&vma->anon_vma_chain);
1704
1705         if (file) {
1706                 if (vm_flags & VM_DENYWRITE) {
1707                         error = deny_write_access(file);
1708                         if (error)
1709                                 goto free_vma;
1710                 }
1711                 if (vm_flags & VM_SHARED) {
1712                         error = mapping_map_writable(file->f_mapping);
1713                         if (error)
1714                                 goto allow_write_and_free_vma;
1715                 }
1716
1717                 /* ->mmap() can change vma->vm_file, but must guarantee that
1718                  * vma_link() below can deny write-access if VM_DENYWRITE is set
1719                  * and map writably if VM_SHARED is set. This usually means the
1720                  * new file must not have been exposed to user-space, yet.
1721                  */
1722                 vma->vm_file = get_file(file);
1723                 error = call_mmap(file, vma);
1724                 if (error)
1725                         goto unmap_and_free_vma;
1726
1727                 /* Can addr have changed??
1728                  *
1729                  * Answer: Yes, several device drivers can do it in their
1730                  *         f_op->mmap method. -DaveM
1731                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1732                  *      be updated for vma_link()
1733                  */
1734                 WARN_ON_ONCE(addr != vma->vm_start);
1735
1736                 addr = vma->vm_start;
1737                 vm_flags = vma->vm_flags;
1738         } else if (vm_flags & VM_SHARED) {
1739                 error = shmem_zero_setup(vma);
1740                 if (error)
1741                         goto free_vma;
1742         }
1743
1744         vma_link(mm, vma, prev, rb_link, rb_parent);
1745         /* Once vma denies write, undo our temporary denial count */
1746         if (file) {
1747                 if (vm_flags & VM_SHARED)
1748                         mapping_unmap_writable(file->f_mapping);
1749                 if (vm_flags & VM_DENYWRITE)
1750                         allow_write_access(file);
1751         }
1752         file = vma->vm_file;
1753 out:
1754         perf_event_mmap(vma);
1755
1756         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1757         if (vm_flags & VM_LOCKED) {
1758                 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1759                                         vma == get_gate_vma(current->mm)))
1760                         mm->locked_vm += (len >> PAGE_SHIFT);
1761                 else
1762                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1763         }
1764
1765         if (file)
1766                 uprobe_mmap(vma);
1767
1768         /*
1769          * New (or expanded) vma always get soft dirty status.
1770          * Otherwise user-space soft-dirty page tracker won't
1771          * be able to distinguish situation when vma area unmapped,
1772          * then new mapped in-place (which must be aimed as
1773          * a completely new data area).
1774          */
1775         vma->vm_flags |= VM_SOFTDIRTY;
1776
1777         vma_set_page_prot(vma);
1778
1779         return addr;
1780
1781 unmap_and_free_vma:
1782         vma->vm_file = NULL;
1783         fput(file);
1784
1785         /* Undo any partial mapping done by a device driver. */
1786         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1787         charged = 0;
1788         if (vm_flags & VM_SHARED)
1789                 mapping_unmap_writable(file->f_mapping);
1790 allow_write_and_free_vma:
1791         if (vm_flags & VM_DENYWRITE)
1792                 allow_write_access(file);
1793 free_vma:
1794         kmem_cache_free(vm_area_cachep, vma);
1795 unacct_error:
1796         if (charged)
1797                 vm_unacct_memory(charged);
1798         return error;
1799 }
1800
1801 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1802 {
1803         /*
1804          * We implement the search by looking for an rbtree node that
1805          * immediately follows a suitable gap. That is,
1806          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1807          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1808          * - gap_end - gap_start >= length
1809          */
1810
1811         struct mm_struct *mm = current->mm;
1812         struct vm_area_struct *vma;
1813         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1814
1815         /* Adjust search length to account for worst case alignment overhead */
1816         length = info->length + info->align_mask;
1817         if (length < info->length)
1818                 return -ENOMEM;
1819
1820         /* Adjust search limits by the desired length */
1821         if (info->high_limit < length)
1822                 return -ENOMEM;
1823         high_limit = info->high_limit - length;
1824
1825         if (info->low_limit > high_limit)
1826                 return -ENOMEM;
1827         low_limit = info->low_limit + length;
1828
1829         /* Check if rbtree root looks promising */
1830         if (RB_EMPTY_ROOT(&mm->mm_rb))
1831                 goto check_highest;
1832         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1833         if (vma->rb_subtree_gap < length)
1834                 goto check_highest;
1835
1836         while (true) {
1837                 /* Visit left subtree if it looks promising */
1838                 gap_end = vm_start_gap(vma);
1839                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1840                         struct vm_area_struct *left =
1841                                 rb_entry(vma->vm_rb.rb_left,
1842                                          struct vm_area_struct, vm_rb);
1843                         if (left->rb_subtree_gap >= length) {
1844                                 vma = left;
1845                                 continue;
1846                         }
1847                 }
1848
1849                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1850 check_current:
1851                 /* Check if current node has a suitable gap */
1852                 if (gap_start > high_limit)
1853                         return -ENOMEM;
1854                 if (gap_end >= low_limit &&
1855                     gap_end > gap_start && gap_end - gap_start >= length)
1856                         goto found;
1857
1858                 /* Visit right subtree if it looks promising */
1859                 if (vma->vm_rb.rb_right) {
1860                         struct vm_area_struct *right =
1861                                 rb_entry(vma->vm_rb.rb_right,
1862                                          struct vm_area_struct, vm_rb);
1863                         if (right->rb_subtree_gap >= length) {
1864                                 vma = right;
1865                                 continue;
1866                         }
1867                 }
1868
1869                 /* Go back up the rbtree to find next candidate node */
1870                 while (true) {
1871                         struct rb_node *prev = &vma->vm_rb;
1872                         if (!rb_parent(prev))
1873                                 goto check_highest;
1874                         vma = rb_entry(rb_parent(prev),
1875                                        struct vm_area_struct, vm_rb);
1876                         if (prev == vma->vm_rb.rb_left) {
1877                                 gap_start = vm_end_gap(vma->vm_prev);
1878                                 gap_end = vm_start_gap(vma);
1879                                 goto check_current;
1880                         }
1881                 }
1882         }
1883
1884 check_highest:
1885         /* Check highest gap, which does not precede any rbtree node */
1886         gap_start = mm->highest_vm_end;
1887         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1888         if (gap_start > high_limit)
1889                 return -ENOMEM;
1890
1891 found:
1892         /* We found a suitable gap. Clip it with the original low_limit. */
1893         if (gap_start < info->low_limit)
1894                 gap_start = info->low_limit;
1895
1896         /* Adjust gap address to the desired alignment */
1897         gap_start += (info->align_offset - gap_start) & info->align_mask;
1898
1899         VM_BUG_ON(gap_start + info->length > info->high_limit);
1900         VM_BUG_ON(gap_start + info->length > gap_end);
1901         return gap_start;
1902 }
1903
1904 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1905 {
1906         struct mm_struct *mm = current->mm;
1907         struct vm_area_struct *vma;
1908         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1909
1910         /* Adjust search length to account for worst case alignment overhead */
1911         length = info->length + info->align_mask;
1912         if (length < info->length)
1913                 return -ENOMEM;
1914
1915         /*
1916          * Adjust search limits by the desired length.
1917          * See implementation comment at top of unmapped_area().
1918          */
1919         gap_end = info->high_limit;
1920         if (gap_end < length)
1921                 return -ENOMEM;
1922         high_limit = gap_end - length;
1923
1924         if (info->low_limit > high_limit)
1925                 return -ENOMEM;
1926         low_limit = info->low_limit + length;
1927
1928         /* Check highest gap, which does not precede any rbtree node */
1929         gap_start = mm->highest_vm_end;
1930         if (gap_start <= high_limit)
1931                 goto found_highest;
1932
1933         /* Check if rbtree root looks promising */
1934         if (RB_EMPTY_ROOT(&mm->mm_rb))
1935                 return -ENOMEM;
1936         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1937         if (vma->rb_subtree_gap < length)
1938                 return -ENOMEM;
1939
1940         while (true) {
1941                 /* Visit right subtree if it looks promising */
1942                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1943                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1944                         struct vm_area_struct *right =
1945                                 rb_entry(vma->vm_rb.rb_right,
1946                                          struct vm_area_struct, vm_rb);
1947                         if (right->rb_subtree_gap >= length) {
1948                                 vma = right;
1949                                 continue;
1950                         }
1951                 }
1952
1953 check_current:
1954                 /* Check if current node has a suitable gap */
1955                 gap_end = vm_start_gap(vma);
1956                 if (gap_end < low_limit)
1957                         return -ENOMEM;
1958                 if (gap_start <= high_limit &&
1959                     gap_end > gap_start && gap_end - gap_start >= length)
1960                         goto found;
1961
1962                 /* Visit left subtree if it looks promising */
1963                 if (vma->vm_rb.rb_left) {
1964                         struct vm_area_struct *left =
1965                                 rb_entry(vma->vm_rb.rb_left,
1966                                          struct vm_area_struct, vm_rb);
1967                         if (left->rb_subtree_gap >= length) {
1968                                 vma = left;
1969                                 continue;
1970                         }
1971                 }
1972
1973                 /* Go back up the rbtree to find next candidate node */
1974                 while (true) {
1975                         struct rb_node *prev = &vma->vm_rb;
1976                         if (!rb_parent(prev))
1977                                 return -ENOMEM;
1978                         vma = rb_entry(rb_parent(prev),
1979                                        struct vm_area_struct, vm_rb);
1980                         if (prev == vma->vm_rb.rb_right) {
1981                                 gap_start = vma->vm_prev ?
1982                                         vm_end_gap(vma->vm_prev) : 0;
1983                                 goto check_current;
1984                         }
1985                 }
1986         }
1987
1988 found:
1989         /* We found a suitable gap. Clip it with the original high_limit. */
1990         if (gap_end > info->high_limit)
1991                 gap_end = info->high_limit;
1992
1993 found_highest:
1994         /* Compute highest gap address at the desired alignment */
1995         gap_end -= info->length;
1996         gap_end -= (gap_end - info->align_offset) & info->align_mask;
1997
1998         VM_BUG_ON(gap_end < info->low_limit);
1999         VM_BUG_ON(gap_end < gap_start);
2000         return gap_end;
2001 }
2002
2003 /* Get an address range which is currently unmapped.
2004  * For shmat() with addr=0.
2005  *
2006  * Ugly calling convention alert:
2007  * Return value with the low bits set means error value,
2008  * ie
2009  *      if (ret & ~PAGE_MASK)
2010  *              error = ret;
2011  *
2012  * This function "knows" that -ENOMEM has the bits set.
2013  */
2014 #ifndef HAVE_ARCH_UNMAPPED_AREA
2015 unsigned long
2016 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2017                 unsigned long len, unsigned long pgoff, unsigned long flags)
2018 {
2019         struct mm_struct *mm = current->mm;
2020         struct vm_area_struct *vma, *prev;
2021         struct vm_unmapped_area_info info;
2022
2023         if (len > TASK_SIZE - mmap_min_addr)
2024                 return -ENOMEM;
2025
2026         if (flags & MAP_FIXED)
2027                 return addr;
2028
2029         if (addr) {
2030                 addr = PAGE_ALIGN(addr);
2031                 vma = find_vma_prev(mm, addr, &prev);
2032                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2033                     (!vma || addr + len <= vm_start_gap(vma)) &&
2034                     (!prev || addr >= vm_end_gap(prev)))
2035                         return addr;
2036         }
2037
2038         info.flags = 0;
2039         info.length = len;
2040         info.low_limit = mm->mmap_base;
2041         info.high_limit = TASK_SIZE;
2042         info.align_mask = 0;
2043         return vm_unmapped_area(&info);
2044 }
2045 #endif
2046
2047 /*
2048  * This mmap-allocator allocates new areas top-down from below the
2049  * stack's low limit (the base):
2050  */
2051 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2052 unsigned long
2053 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2054                           const unsigned long len, const unsigned long pgoff,
2055                           const unsigned long flags)
2056 {
2057         struct vm_area_struct *vma, *prev;
2058         struct mm_struct *mm = current->mm;
2059         unsigned long addr = addr0;
2060         struct vm_unmapped_area_info info;
2061
2062         /* requested length too big for entire address space */
2063         if (len > TASK_SIZE - mmap_min_addr)
2064                 return -ENOMEM;
2065
2066         if (flags & MAP_FIXED)
2067                 return addr;
2068
2069         /* requesting a specific address */
2070         if (addr) {
2071                 addr = PAGE_ALIGN(addr);
2072                 vma = find_vma_prev(mm, addr, &prev);
2073                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2074                                 (!vma || addr + len <= vm_start_gap(vma)) &&
2075                                 (!prev || addr >= vm_end_gap(prev)))
2076                         return addr;
2077         }
2078
2079         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2080         info.length = len;
2081         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2082         info.high_limit = mm->mmap_base;
2083         info.align_mask = 0;
2084         addr = vm_unmapped_area(&info);
2085
2086         /*
2087          * A failed mmap() very likely causes application failure,
2088          * so fall back to the bottom-up function here. This scenario
2089          * can happen with large stack limits and large mmap()
2090          * allocations.
2091          */
2092         if (offset_in_page(addr)) {
2093                 VM_BUG_ON(addr != -ENOMEM);
2094                 info.flags = 0;
2095                 info.low_limit = TASK_UNMAPPED_BASE;
2096                 info.high_limit = TASK_SIZE;
2097                 addr = vm_unmapped_area(&info);
2098         }
2099
2100         return addr;
2101 }
2102 #endif
2103
2104 unsigned long
2105 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2106                 unsigned long pgoff, unsigned long flags)
2107 {
2108         unsigned long (*get_area)(struct file *, unsigned long,
2109                                   unsigned long, unsigned long, unsigned long);
2110
2111         unsigned long error = arch_mmap_check(addr, len, flags);
2112         if (error)
2113                 return error;
2114
2115         /* Careful about overflows.. */
2116         if (len > TASK_SIZE)
2117                 return -ENOMEM;
2118
2119         get_area = current->mm->get_unmapped_area;
2120         if (file) {
2121                 if (file->f_op->get_unmapped_area)
2122                         get_area = file->f_op->get_unmapped_area;
2123         } else if (flags & MAP_SHARED) {
2124                 /*
2125                  * mmap_region() will call shmem_zero_setup() to create a file,
2126                  * so use shmem's get_unmapped_area in case it can be huge.
2127                  * do_mmap_pgoff() will clear pgoff, so match alignment.
2128                  */
2129                 pgoff = 0;
2130                 get_area = shmem_get_unmapped_area;
2131         }
2132
2133         addr = get_area(file, addr, len, pgoff, flags);
2134         if (IS_ERR_VALUE(addr))
2135                 return addr;
2136
2137         if (addr > TASK_SIZE - len)
2138                 return -ENOMEM;
2139         if (offset_in_page(addr))
2140                 return -EINVAL;
2141
2142         error = security_mmap_addr(addr);
2143         return error ? error : addr;
2144 }
2145
2146 EXPORT_SYMBOL(get_unmapped_area);
2147
2148 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2149 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2150 {
2151         struct rb_node *rb_node;
2152         struct vm_area_struct *vma;
2153
2154         /* Check the cache first. */
2155         vma = vmacache_find(mm, addr);
2156         if (likely(vma))
2157                 return vma;
2158
2159         rb_node = mm->mm_rb.rb_node;
2160
2161         while (rb_node) {
2162                 struct vm_area_struct *tmp;
2163
2164                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2165
2166                 if (tmp->vm_end > addr) {
2167                         vma = tmp;
2168                         if (tmp->vm_start <= addr)
2169                                 break;
2170                         rb_node = rb_node->rb_left;
2171                 } else
2172                         rb_node = rb_node->rb_right;
2173         }
2174
2175         if (vma)
2176                 vmacache_update(addr, vma);
2177         return vma;
2178 }
2179
2180 EXPORT_SYMBOL(find_vma);
2181
2182 /*
2183  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2184  */
2185 struct vm_area_struct *
2186 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2187                         struct vm_area_struct **pprev)
2188 {
2189         struct vm_area_struct *vma;
2190
2191         vma = find_vma(mm, addr);
2192         if (vma) {
2193                 *pprev = vma->vm_prev;
2194         } else {
2195                 struct rb_node *rb_node = mm->mm_rb.rb_node;
2196                 *pprev = NULL;
2197                 while (rb_node) {
2198                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2199                         rb_node = rb_node->rb_right;
2200                 }
2201         }
2202         return vma;
2203 }
2204
2205 /*
2206  * Verify that the stack growth is acceptable and
2207  * update accounting. This is shared with both the
2208  * grow-up and grow-down cases.
2209  */
2210 static int acct_stack_growth(struct vm_area_struct *vma,
2211                              unsigned long size, unsigned long grow)
2212 {
2213         struct mm_struct *mm = vma->vm_mm;
2214         unsigned long new_start;
2215
2216         /* address space limit tests */
2217         if (!may_expand_vm(mm, vma->vm_flags, grow))
2218                 return -ENOMEM;
2219
2220         /* Stack limit test */
2221         if (size > rlimit(RLIMIT_STACK))
2222                 return -ENOMEM;
2223
2224         /* mlock limit tests */
2225         if (vma->vm_flags & VM_LOCKED) {
2226                 unsigned long locked;
2227                 unsigned long limit;
2228                 locked = mm->locked_vm + grow;
2229                 limit = rlimit(RLIMIT_MEMLOCK);
2230                 limit >>= PAGE_SHIFT;
2231                 if (locked > limit && !capable(CAP_IPC_LOCK))
2232                         return -ENOMEM;
2233         }
2234
2235         /* Check to ensure the stack will not grow into a hugetlb-only region */
2236         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2237                         vma->vm_end - size;
2238         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2239                 return -EFAULT;
2240
2241         /*
2242          * Overcommit..  This must be the final test, as it will
2243          * update security statistics.
2244          */
2245         if (security_vm_enough_memory_mm(mm, grow))
2246                 return -ENOMEM;
2247
2248         return 0;
2249 }
2250
2251 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2252 /*
2253  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2254  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2255  */
2256 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2257 {
2258         struct mm_struct *mm = vma->vm_mm;
2259         struct vm_area_struct *next;
2260         unsigned long gap_addr;
2261         int error = 0;
2262
2263         if (!(vma->vm_flags & VM_GROWSUP))
2264                 return -EFAULT;
2265
2266         /* Guard against exceeding limits of the address space. */
2267         address &= PAGE_MASK;
2268         if (address >= (TASK_SIZE & PAGE_MASK))
2269                 return -ENOMEM;
2270         address += PAGE_SIZE;
2271
2272         /* Enforce stack_guard_gap */
2273         gap_addr = address + stack_guard_gap;
2274
2275         /* Guard against overflow */
2276         if (gap_addr < address || gap_addr > TASK_SIZE)
2277                 gap_addr = TASK_SIZE;
2278
2279         next = vma->vm_next;
2280         if (next && next->vm_start < gap_addr &&
2281                         (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2282                 if (!(next->vm_flags & VM_GROWSUP))
2283                         return -ENOMEM;
2284                 /* Check that both stack segments have the same anon_vma? */
2285         }
2286
2287         /* We must make sure the anon_vma is allocated. */
2288         if (unlikely(anon_vma_prepare(vma)))
2289                 return -ENOMEM;
2290
2291         /*
2292          * vma->vm_start/vm_end cannot change under us because the caller
2293          * is required to hold the mmap_sem in read mode.  We need the
2294          * anon_vma lock to serialize against concurrent expand_stacks.
2295          */
2296         anon_vma_lock_write(vma->anon_vma);
2297
2298         /* Somebody else might have raced and expanded it already */
2299         if (address > vma->vm_end) {
2300                 unsigned long size, grow;
2301
2302                 size = address - vma->vm_start;
2303                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2304
2305                 error = -ENOMEM;
2306                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2307                         error = acct_stack_growth(vma, size, grow);
2308                         if (!error) {
2309                                 /*
2310                                  * vma_gap_update() doesn't support concurrent
2311                                  * updates, but we only hold a shared mmap_sem
2312                                  * lock here, so we need to protect against
2313                                  * concurrent vma expansions.
2314                                  * anon_vma_lock_write() doesn't help here, as
2315                                  * we don't guarantee that all growable vmas
2316                                  * in a mm share the same root anon vma.
2317                                  * So, we reuse mm->page_table_lock to guard
2318                                  * against concurrent vma expansions.
2319                                  */
2320                                 spin_lock(&mm->page_table_lock);
2321                                 if (vma->vm_flags & VM_LOCKED)
2322                                         mm->locked_vm += grow;
2323                                 vm_stat_account(mm, vma->vm_flags, grow);
2324                                 anon_vma_interval_tree_pre_update_vma(vma);
2325                                 vma->vm_end = address;
2326                                 anon_vma_interval_tree_post_update_vma(vma);
2327                                 if (vma->vm_next)
2328                                         vma_gap_update(vma->vm_next);
2329                                 else
2330                                         mm->highest_vm_end = vm_end_gap(vma);
2331                                 spin_unlock(&mm->page_table_lock);
2332
2333                                 perf_event_mmap(vma);
2334                         }
2335                 }
2336         }
2337         anon_vma_unlock_write(vma->anon_vma);
2338         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2339         validate_mm(mm);
2340         return error;
2341 }
2342 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2343
2344 /*
2345  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2346  */
2347 int expand_downwards(struct vm_area_struct *vma,
2348                                    unsigned long address)
2349 {
2350         struct mm_struct *mm = vma->vm_mm;
2351         struct vm_area_struct *prev;
2352         int error;
2353
2354         address &= PAGE_MASK;
2355         error = security_mmap_addr(address);
2356         if (error)
2357                 return error;
2358
2359         /* Enforce stack_guard_gap */
2360         prev = vma->vm_prev;
2361         /* Check that both stack segments have the same anon_vma? */
2362         if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2363                         (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2364                 if (address - prev->vm_end < stack_guard_gap)
2365                         return -ENOMEM;
2366         }
2367
2368         /* We must make sure the anon_vma is allocated. */
2369         if (unlikely(anon_vma_prepare(vma)))
2370                 return -ENOMEM;
2371
2372         /*
2373          * vma->vm_start/vm_end cannot change under us because the caller
2374          * is required to hold the mmap_sem in read mode.  We need the
2375          * anon_vma lock to serialize against concurrent expand_stacks.
2376          */
2377         anon_vma_lock_write(vma->anon_vma);
2378
2379         /* Somebody else might have raced and expanded it already */
2380         if (address < vma->vm_start) {
2381                 unsigned long size, grow;
2382
2383                 size = vma->vm_end - address;
2384                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2385
2386                 error = -ENOMEM;
2387                 if (grow <= vma->vm_pgoff) {
2388                         error = acct_stack_growth(vma, size, grow);
2389                         if (!error) {
2390                                 /*
2391                                  * vma_gap_update() doesn't support concurrent
2392                                  * updates, but we only hold a shared mmap_sem
2393                                  * lock here, so we need to protect against
2394                                  * concurrent vma expansions.
2395                                  * anon_vma_lock_write() doesn't help here, as
2396                                  * we don't guarantee that all growable vmas
2397                                  * in a mm share the same root anon vma.
2398                                  * So, we reuse mm->page_table_lock to guard
2399                                  * against concurrent vma expansions.
2400                                  */
2401                                 spin_lock(&mm->page_table_lock);
2402                                 if (vma->vm_flags & VM_LOCKED)
2403                                         mm->locked_vm += grow;
2404                                 vm_stat_account(mm, vma->vm_flags, grow);
2405                                 anon_vma_interval_tree_pre_update_vma(vma);
2406                                 vma->vm_start = address;
2407                                 vma->vm_pgoff -= grow;
2408                                 anon_vma_interval_tree_post_update_vma(vma);
2409                                 vma_gap_update(vma);
2410                                 spin_unlock(&mm->page_table_lock);
2411
2412                                 perf_event_mmap(vma);
2413                         }
2414                 }
2415         }
2416         anon_vma_unlock_write(vma->anon_vma);
2417         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2418         validate_mm(mm);
2419         return error;
2420 }
2421
2422 /* enforced gap between the expanding stack and other mappings. */
2423 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2424
2425 static int __init cmdline_parse_stack_guard_gap(char *p)
2426 {
2427         unsigned long val;
2428         char *endptr;
2429
2430         val = simple_strtoul(p, &endptr, 10);
2431         if (!*endptr)
2432                 stack_guard_gap = val << PAGE_SHIFT;
2433
2434         return 0;
2435 }
2436 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2437
2438 #ifdef CONFIG_STACK_GROWSUP
2439 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2440 {
2441         return expand_upwards(vma, address);
2442 }
2443
2444 struct vm_area_struct *
2445 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2446 {
2447         struct vm_area_struct *vma, *prev;
2448
2449         addr &= PAGE_MASK;
2450         vma = find_vma_prev(mm, addr, &prev);
2451         if (vma && (vma->vm_start <= addr))
2452                 return vma;
2453         if (!prev || expand_stack(prev, addr))
2454                 return NULL;
2455         if (prev->vm_flags & VM_LOCKED)
2456                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2457         return prev;
2458 }
2459 #else
2460 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2461 {
2462         return expand_downwards(vma, address);
2463 }
2464
2465 struct vm_area_struct *
2466 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2467 {
2468         struct vm_area_struct *vma;
2469         unsigned long start;
2470
2471         addr &= PAGE_MASK;
2472         vma = find_vma(mm, addr);
2473         if (!vma)
2474                 return NULL;
2475         if (vma->vm_start <= addr)
2476                 return vma;
2477         if (!(vma->vm_flags & VM_GROWSDOWN))
2478                 return NULL;
2479         start = vma->vm_start;
2480         if (expand_stack(vma, addr))
2481                 return NULL;
2482         if (vma->vm_flags & VM_LOCKED)
2483                 populate_vma_page_range(vma, addr, start, NULL);
2484         return vma;
2485 }
2486 #endif
2487
2488 EXPORT_SYMBOL_GPL(find_extend_vma);
2489
2490 /*
2491  * Ok - we have the memory areas we should free on the vma list,
2492  * so release them, and do the vma updates.
2493  *
2494  * Called with the mm semaphore held.
2495  */
2496 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2497 {
2498         unsigned long nr_accounted = 0;
2499
2500         /* Update high watermark before we lower total_vm */
2501         update_hiwater_vm(mm);
2502         do {
2503                 long nrpages = vma_pages(vma);
2504
2505                 if (vma->vm_flags & VM_ACCOUNT)
2506                         nr_accounted += nrpages;
2507                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2508                 vma = remove_vma(vma);
2509         } while (vma);
2510         vm_unacct_memory(nr_accounted);
2511         validate_mm(mm);
2512 }
2513
2514 /*
2515  * Get rid of page table information in the indicated region.
2516  *
2517  * Called with the mm semaphore held.
2518  */
2519 static void unmap_region(struct mm_struct *mm,
2520                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2521                 unsigned long start, unsigned long end)
2522 {
2523         struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2524         struct mmu_gather tlb;
2525
2526         lru_add_drain();
2527         tlb_gather_mmu(&tlb, mm, start, end);
2528         update_hiwater_rss(mm);
2529         unmap_vmas(&tlb, vma, start, end);
2530         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2531                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2532         tlb_finish_mmu(&tlb, start, end);
2533 }
2534
2535 /*
2536  * Create a list of vma's touched by the unmap, removing them from the mm's
2537  * vma list as we go..
2538  */
2539 static void
2540 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2541         struct vm_area_struct *prev, unsigned long end)
2542 {
2543         struct vm_area_struct **insertion_point;
2544         struct vm_area_struct *tail_vma = NULL;
2545
2546         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2547         vma->vm_prev = NULL;
2548         do {
2549                 vma_rb_erase(vma, &mm->mm_rb);
2550                 mm->map_count--;
2551                 tail_vma = vma;
2552                 vma = vma->vm_next;
2553         } while (vma && vma->vm_start < end);
2554         *insertion_point = vma;
2555         if (vma) {
2556                 vma->vm_prev = prev;
2557                 vma_gap_update(vma);
2558         } else
2559                 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2560         tail_vma->vm_next = NULL;
2561
2562         /* Kill the cache */
2563         vmacache_invalidate(mm);
2564 }
2565
2566 /*
2567  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2568  * has already been checked or doesn't make sense to fail.
2569  */
2570 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2571                 unsigned long addr, int new_below)
2572 {
2573         struct vm_area_struct *new;
2574         int err;
2575
2576         if (vma->vm_ops && vma->vm_ops->split) {
2577                 err = vma->vm_ops->split(vma, addr);
2578                 if (err)
2579                         return err;
2580         }
2581
2582         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2583         if (!new)
2584                 return -ENOMEM;
2585
2586         /* most fields are the same, copy all, and then fixup */
2587         *new = *vma;
2588
2589         INIT_LIST_HEAD(&new->anon_vma_chain);
2590
2591         if (new_below)
2592                 new->vm_end = addr;
2593         else {
2594                 new->vm_start = addr;
2595                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2596         }
2597
2598         err = vma_dup_policy(vma, new);
2599         if (err)
2600                 goto out_free_vma;
2601
2602         err = anon_vma_clone(new, vma);
2603         if (err)
2604                 goto out_free_mpol;
2605
2606         if (new->vm_file)
2607                 get_file(new->vm_file);
2608
2609         if (new->vm_ops && new->vm_ops->open)
2610                 new->vm_ops->open(new);
2611
2612         if (new_below)
2613                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2614                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2615         else
2616                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2617
2618         /* Success. */
2619         if (!err)
2620                 return 0;
2621
2622         /* Clean everything up if vma_adjust failed. */
2623         if (new->vm_ops && new->vm_ops->close)
2624                 new->vm_ops->close(new);
2625         if (new->vm_file)
2626                 fput(new->vm_file);
2627         unlink_anon_vmas(new);
2628  out_free_mpol:
2629         mpol_put(vma_policy(new));
2630  out_free_vma:
2631         kmem_cache_free(vm_area_cachep, new);
2632         return err;
2633 }
2634
2635 /*
2636  * Split a vma into two pieces at address 'addr', a new vma is allocated
2637  * either for the first part or the tail.
2638  */
2639 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2640               unsigned long addr, int new_below)
2641 {
2642         if (mm->map_count >= sysctl_max_map_count)
2643                 return -ENOMEM;
2644
2645         return __split_vma(mm, vma, addr, new_below);
2646 }
2647
2648 /* Munmap is split into 2 main parts -- this part which finds
2649  * what needs doing, and the areas themselves, which do the
2650  * work.  This now handles partial unmappings.
2651  * Jeremy Fitzhardinge <jeremy@goop.org>
2652  */
2653 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2654               struct list_head *uf)
2655 {
2656         unsigned long end;
2657         struct vm_area_struct *vma, *prev, *last;
2658
2659         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2660                 return -EINVAL;
2661
2662         len = PAGE_ALIGN(len);
2663         if (len == 0)
2664                 return -EINVAL;
2665
2666         /* Find the first overlapping VMA */
2667         vma = find_vma(mm, start);
2668         if (!vma)
2669                 return 0;
2670         prev = vma->vm_prev;
2671         /* we have  start < vma->vm_end  */
2672
2673         /* if it doesn't overlap, we have nothing.. */
2674         end = start + len;
2675         if (vma->vm_start >= end)
2676                 return 0;
2677
2678         /*
2679          * If we need to split any vma, do it now to save pain later.
2680          *
2681          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2682          * unmapped vm_area_struct will remain in use: so lower split_vma
2683          * places tmp vma above, and higher split_vma places tmp vma below.
2684          */
2685         if (start > vma->vm_start) {
2686                 int error;
2687
2688                 /*
2689                  * Make sure that map_count on return from munmap() will
2690                  * not exceed its limit; but let map_count go just above
2691                  * its limit temporarily, to help free resources as expected.
2692                  */
2693                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2694                         return -ENOMEM;
2695
2696                 error = __split_vma(mm, vma, start, 0);
2697                 if (error)
2698                         return error;
2699                 prev = vma;
2700         }
2701
2702         /* Does it split the last one? */
2703         last = find_vma(mm, end);
2704         if (last && end > last->vm_start) {
2705                 int error = __split_vma(mm, last, end, 1);
2706                 if (error)
2707                         return error;
2708         }
2709         vma = prev ? prev->vm_next : mm->mmap;
2710
2711         if (unlikely(uf)) {
2712                 /*
2713                  * If userfaultfd_unmap_prep returns an error the vmas
2714                  * will remain splitted, but userland will get a
2715                  * highly unexpected error anyway. This is no
2716                  * different than the case where the first of the two
2717                  * __split_vma fails, but we don't undo the first
2718                  * split, despite we could. This is unlikely enough
2719                  * failure that it's not worth optimizing it for.
2720                  */
2721                 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2722                 if (error)
2723                         return error;
2724         }
2725
2726         /*
2727          * unlock any mlock()ed ranges before detaching vmas
2728          */
2729         if (mm->locked_vm) {
2730                 struct vm_area_struct *tmp = vma;
2731                 while (tmp && tmp->vm_start < end) {
2732                         if (tmp->vm_flags & VM_LOCKED) {
2733                                 mm->locked_vm -= vma_pages(tmp);
2734                                 munlock_vma_pages_all(tmp);
2735                         }
2736                         tmp = tmp->vm_next;
2737                 }
2738         }
2739
2740         /*
2741          * Remove the vma's, and unmap the actual pages
2742          */
2743         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2744         unmap_region(mm, vma, prev, start, end);
2745
2746         arch_unmap(mm, vma, start, end);
2747
2748         /* Fix up all other VM information */
2749         remove_vma_list(mm, vma);
2750
2751         return 0;
2752 }
2753
2754 int vm_munmap(unsigned long start, size_t len)
2755 {
2756         int ret;
2757         struct mm_struct *mm = current->mm;
2758         LIST_HEAD(uf);
2759
2760         if (down_write_killable(&mm->mmap_sem))
2761                 return -EINTR;
2762
2763         ret = do_munmap(mm, start, len, &uf);
2764         up_write(&mm->mmap_sem);
2765         userfaultfd_unmap_complete(mm, &uf);
2766         return ret;
2767 }
2768 EXPORT_SYMBOL(vm_munmap);
2769
2770 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2771 {
2772         profile_munmap(addr);
2773         return vm_munmap(addr, len);
2774 }
2775
2776
2777 /*
2778  * Emulation of deprecated remap_file_pages() syscall.
2779  */
2780 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2781                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2782 {
2783
2784         struct mm_struct *mm = current->mm;
2785         struct vm_area_struct *vma;
2786         unsigned long populate = 0;
2787         unsigned long ret = -EINVAL;
2788         struct file *file;
2789
2790         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2791                      current->comm, current->pid);
2792
2793         if (prot)
2794                 return ret;
2795         start = start & PAGE_MASK;
2796         size = size & PAGE_MASK;
2797
2798         if (start + size <= start)
2799                 return ret;
2800
2801         /* Does pgoff wrap? */
2802         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2803                 return ret;
2804
2805         if (down_write_killable(&mm->mmap_sem))
2806                 return -EINTR;
2807
2808         vma = find_vma(mm, start);
2809
2810         if (!vma || !(vma->vm_flags & VM_SHARED))
2811                 goto out;
2812
2813         if (start < vma->vm_start)
2814                 goto out;
2815
2816         if (start + size > vma->vm_end) {
2817                 struct vm_area_struct *next;
2818
2819                 for (next = vma->vm_next; next; next = next->vm_next) {
2820                         /* hole between vmas ? */
2821                         if (next->vm_start != next->vm_prev->vm_end)
2822                                 goto out;
2823
2824                         if (next->vm_file != vma->vm_file)
2825                                 goto out;
2826
2827                         if (next->vm_flags != vma->vm_flags)
2828                                 goto out;
2829
2830                         if (start + size <= next->vm_end)
2831                                 break;
2832                 }
2833
2834                 if (!next)
2835                         goto out;
2836         }
2837
2838         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2839         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2840         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2841
2842         flags &= MAP_NONBLOCK;
2843         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2844         if (vma->vm_flags & VM_LOCKED) {
2845                 struct vm_area_struct *tmp;
2846                 flags |= MAP_LOCKED;
2847
2848                 /* drop PG_Mlocked flag for over-mapped range */
2849                 for (tmp = vma; tmp->vm_start >= start + size;
2850                                 tmp = tmp->vm_next) {
2851                         /*
2852                          * Split pmd and munlock page on the border
2853                          * of the range.
2854                          */
2855                         vma_adjust_trans_huge(tmp, start, start + size, 0);
2856
2857                         munlock_vma_pages_range(tmp,
2858                                         max(tmp->vm_start, start),
2859                                         min(tmp->vm_end, start + size));
2860                 }
2861         }
2862
2863         file = get_file(vma->vm_file);
2864         ret = do_mmap_pgoff(vma->vm_file, start, size,
2865                         prot, flags, pgoff, &populate, NULL);
2866         fput(file);
2867 out:
2868         up_write(&mm->mmap_sem);
2869         if (populate)
2870                 mm_populate(ret, populate);
2871         if (!IS_ERR_VALUE(ret))
2872                 ret = 0;
2873         return ret;
2874 }
2875
2876 static inline void verify_mm_writelocked(struct mm_struct *mm)
2877 {
2878 #ifdef CONFIG_DEBUG_VM
2879         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2880                 WARN_ON(1);
2881                 up_read(&mm->mmap_sem);
2882         }
2883 #endif
2884 }
2885
2886 /*
2887  *  this is really a simplified "do_mmap".  it only handles
2888  *  anonymous maps.  eventually we may be able to do some
2889  *  brk-specific accounting here.
2890  */
2891 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags, struct list_head *uf)
2892 {
2893         struct mm_struct *mm = current->mm;
2894         struct vm_area_struct *vma, *prev;
2895         unsigned long len;
2896         struct rb_node **rb_link, *rb_parent;
2897         pgoff_t pgoff = addr >> PAGE_SHIFT;
2898         int error;
2899
2900         len = PAGE_ALIGN(request);
2901         if (len < request)
2902                 return -ENOMEM;
2903         if (!len)
2904                 return 0;
2905
2906         /* Until we need other flags, refuse anything except VM_EXEC. */
2907         if ((flags & (~VM_EXEC)) != 0)
2908                 return -EINVAL;
2909         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2910
2911         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2912         if (offset_in_page(error))
2913                 return error;
2914
2915         error = mlock_future_check(mm, mm->def_flags, len);
2916         if (error)
2917                 return error;
2918
2919         /*
2920          * mm->mmap_sem is required to protect against another thread
2921          * changing the mappings in case we sleep.
2922          */
2923         verify_mm_writelocked(mm);
2924
2925         /*
2926          * Clear old maps.  this also does some error checking for us
2927          */
2928         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2929                               &rb_parent)) {
2930                 if (do_munmap(mm, addr, len, uf))
2931                         return -ENOMEM;
2932         }
2933
2934         /* Check against address space limits *after* clearing old maps... */
2935         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2936                 return -ENOMEM;
2937
2938         if (mm->map_count > sysctl_max_map_count)
2939                 return -ENOMEM;
2940
2941         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2942                 return -ENOMEM;
2943
2944         /* Can we just expand an old private anonymous mapping? */
2945         vma = vma_merge(mm, prev, addr, addr + len, flags,
2946                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2947         if (vma)
2948                 goto out;
2949
2950         /*
2951          * create a vma struct for an anonymous mapping
2952          */
2953         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2954         if (!vma) {
2955                 vm_unacct_memory(len >> PAGE_SHIFT);
2956                 return -ENOMEM;
2957         }
2958
2959         INIT_LIST_HEAD(&vma->anon_vma_chain);
2960         vma->vm_mm = mm;
2961         vma->vm_start = addr;
2962         vma->vm_end = addr + len;
2963         vma->vm_pgoff = pgoff;
2964         vma->vm_flags = flags;
2965         vma->vm_page_prot = vm_get_page_prot(flags);
2966         vma_link(mm, vma, prev, rb_link, rb_parent);
2967 out:
2968         perf_event_mmap(vma);
2969         mm->total_vm += len >> PAGE_SHIFT;
2970         mm->data_vm += len >> PAGE_SHIFT;
2971         if (flags & VM_LOCKED)
2972                 mm->locked_vm += (len >> PAGE_SHIFT);
2973         vma->vm_flags |= VM_SOFTDIRTY;
2974         return 0;
2975 }
2976
2977 static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf)
2978 {
2979         return do_brk_flags(addr, len, 0, uf);
2980 }
2981
2982 int vm_brk_flags(unsigned long addr, unsigned long len, unsigned long flags)
2983 {
2984         struct mm_struct *mm = current->mm;
2985         int ret;
2986         bool populate;
2987         LIST_HEAD(uf);
2988
2989         if (down_write_killable(&mm->mmap_sem))
2990                 return -EINTR;
2991
2992         ret = do_brk_flags(addr, len, flags, &uf);
2993         populate = ((mm->def_flags & VM_LOCKED) != 0);
2994         up_write(&mm->mmap_sem);
2995         userfaultfd_unmap_complete(mm, &uf);
2996         if (populate && !ret)
2997                 mm_populate(addr, len);
2998         return ret;
2999 }
3000 EXPORT_SYMBOL(vm_brk_flags);
3001
3002 int vm_brk(unsigned long addr, unsigned long len)
3003 {
3004         return vm_brk_flags(addr, len, 0);
3005 }
3006 EXPORT_SYMBOL(vm_brk);
3007
3008 /* Release all mmaps. */
3009 void exit_mmap(struct mm_struct *mm)
3010 {
3011         struct mmu_gather tlb;
3012         struct vm_area_struct *vma;
3013         unsigned long nr_accounted = 0;
3014
3015         /* mm's last user has gone, and its about to be pulled down */
3016         mmu_notifier_release(mm);
3017
3018         if (mm->locked_vm) {
3019                 vma = mm->mmap;
3020                 while (vma) {
3021                         if (vma->vm_flags & VM_LOCKED)
3022                                 munlock_vma_pages_all(vma);
3023                         vma = vma->vm_next;
3024                 }
3025         }
3026
3027         arch_exit_mmap(mm);
3028
3029         vma = mm->mmap;
3030         if (!vma)       /* Can happen if dup_mmap() received an OOM */
3031                 return;
3032
3033         lru_add_drain();
3034         flush_cache_mm(mm);
3035         tlb_gather_mmu(&tlb, mm, 0, -1);
3036         /* update_hiwater_rss(mm) here? but nobody should be looking */
3037         /* Use -1 here to ensure all VMAs in the mm are unmapped */
3038         unmap_vmas(&tlb, vma, 0, -1);
3039
3040         if (unlikely(mm_is_oom_victim(mm))) {
3041                 /*
3042                  * Wait for oom_reap_task() to stop working on this
3043                  * mm. Because MMF_OOM_SKIP is already set before
3044                  * calling down_read(), oom_reap_task() will not run
3045                  * on this "mm" post up_write().
3046                  *
3047                  * mm_is_oom_victim() cannot be set from under us
3048                  * either because victim->mm is already set to NULL
3049                  * under task_lock before calling mmput and oom_mm is
3050                  * set not NULL by the OOM killer only if victim->mm
3051                  * is found not NULL while holding the task_lock.
3052                  */
3053                 set_bit(MMF_OOM_SKIP, &mm->flags);
3054                 down_write(&mm->mmap_sem);
3055                 up_write(&mm->mmap_sem);
3056         }
3057         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3058         tlb_finish_mmu(&tlb, 0, -1);
3059
3060         /*
3061          * Walk the list again, actually closing and freeing it,
3062          * with preemption enabled, without holding any MM locks.
3063          */
3064         while (vma) {
3065                 if (vma->vm_flags & VM_ACCOUNT)
3066                         nr_accounted += vma_pages(vma);
3067                 vma = remove_vma(vma);
3068         }
3069         vm_unacct_memory(nr_accounted);
3070 }
3071
3072 /* Insert vm structure into process list sorted by address
3073  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3074  * then i_mmap_rwsem is taken here.
3075  */
3076 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3077 {
3078         struct vm_area_struct *prev;
3079         struct rb_node **rb_link, *rb_parent;
3080
3081         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3082                            &prev, &rb_link, &rb_parent))
3083                 return -ENOMEM;
3084         if ((vma->vm_flags & VM_ACCOUNT) &&
3085              security_vm_enough_memory_mm(mm, vma_pages(vma)))
3086                 return -ENOMEM;
3087
3088         /*
3089          * The vm_pgoff of a purely anonymous vma should be irrelevant
3090          * until its first write fault, when page's anon_vma and index
3091          * are set.  But now set the vm_pgoff it will almost certainly
3092          * end up with (unless mremap moves it elsewhere before that
3093          * first wfault), so /proc/pid/maps tells a consistent story.
3094          *
3095          * By setting it to reflect the virtual start address of the
3096          * vma, merges and splits can happen in a seamless way, just
3097          * using the existing file pgoff checks and manipulations.
3098          * Similarly in do_mmap_pgoff and in do_brk.
3099          */
3100         if (vma_is_anonymous(vma)) {
3101                 BUG_ON(vma->anon_vma);
3102                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3103         }
3104
3105         vma_link(mm, vma, prev, rb_link, rb_parent);
3106         return 0;
3107 }
3108
3109 /*
3110  * Copy the vma structure to a new location in the same mm,
3111  * prior to moving page table entries, to effect an mremap move.
3112  */
3113 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3114         unsigned long addr, unsigned long len, pgoff_t pgoff,
3115         bool *need_rmap_locks)
3116 {
3117         struct vm_area_struct *vma = *vmap;
3118         unsigned long vma_start = vma->vm_start;
3119         struct mm_struct *mm = vma->vm_mm;
3120         struct vm_area_struct *new_vma, *prev;
3121         struct rb_node **rb_link, *rb_parent;
3122         bool faulted_in_anon_vma = true;
3123
3124         /*
3125          * If anonymous vma has not yet been faulted, update new pgoff
3126          * to match new location, to increase its chance of merging.
3127          */
3128         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3129                 pgoff = addr >> PAGE_SHIFT;
3130                 faulted_in_anon_vma = false;
3131         }
3132
3133         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3134                 return NULL;    /* should never get here */
3135         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3136                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3137                             vma->vm_userfaultfd_ctx);
3138         if (new_vma) {
3139                 /*
3140                  * Source vma may have been merged into new_vma
3141                  */
3142                 if (unlikely(vma_start >= new_vma->vm_start &&
3143                              vma_start < new_vma->vm_end)) {
3144                         /*
3145                          * The only way we can get a vma_merge with
3146                          * self during an mremap is if the vma hasn't
3147                          * been faulted in yet and we were allowed to
3148                          * reset the dst vma->vm_pgoff to the
3149                          * destination address of the mremap to allow
3150                          * the merge to happen. mremap must change the
3151                          * vm_pgoff linearity between src and dst vmas
3152                          * (in turn preventing a vma_merge) to be
3153                          * safe. It is only safe to keep the vm_pgoff
3154                          * linear if there are no pages mapped yet.
3155                          */
3156                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3157                         *vmap = vma = new_vma;
3158                 }
3159                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3160         } else {
3161                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3162                 if (!new_vma)
3163                         goto out;
3164                 *new_vma = *vma;
3165                 new_vma->vm_start = addr;
3166                 new_vma->vm_end = addr + len;
3167                 new_vma->vm_pgoff = pgoff;
3168                 if (vma_dup_policy(vma, new_vma))
3169                         goto out_free_vma;
3170                 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3171                 if (anon_vma_clone(new_vma, vma))
3172                         goto out_free_mempol;
3173                 if (new_vma->vm_file)
3174                         get_file(new_vma->vm_file);
3175                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3176                         new_vma->vm_ops->open(new_vma);
3177                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3178                 *need_rmap_locks = false;
3179         }
3180         return new_vma;
3181
3182 out_free_mempol:
3183         mpol_put(vma_policy(new_vma));
3184 out_free_vma:
3185         kmem_cache_free(vm_area_cachep, new_vma);
3186 out:
3187         return NULL;
3188 }
3189
3190 /*
3191  * Return true if the calling process may expand its vm space by the passed
3192  * number of pages
3193  */
3194 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3195 {
3196         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3197                 return false;
3198
3199         if (is_data_mapping(flags) &&
3200             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3201                 /* Workaround for Valgrind */
3202                 if (rlimit(RLIMIT_DATA) == 0 &&
3203                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3204                         return true;
3205
3206                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3207                              current->comm, current->pid,
3208                              (mm->data_vm + npages) << PAGE_SHIFT,
3209                              rlimit(RLIMIT_DATA),
3210                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3211
3212                 if (!ignore_rlimit_data)
3213                         return false;
3214         }
3215
3216         return true;
3217 }
3218
3219 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3220 {
3221         mm->total_vm += npages;
3222
3223         if (is_exec_mapping(flags))
3224                 mm->exec_vm += npages;
3225         else if (is_stack_mapping(flags))
3226                 mm->stack_vm += npages;
3227         else if (is_data_mapping(flags))
3228                 mm->data_vm += npages;
3229 }
3230
3231 static int special_mapping_fault(struct vm_fault *vmf);
3232
3233 /*
3234  * Having a close hook prevents vma merging regardless of flags.
3235  */
3236 static void special_mapping_close(struct vm_area_struct *vma)
3237 {
3238 }
3239
3240 static const char *special_mapping_name(struct vm_area_struct *vma)
3241 {
3242         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3243 }
3244
3245 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3246 {
3247         struct vm_special_mapping *sm = new_vma->vm_private_data;
3248
3249         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3250                 return -EFAULT;
3251
3252         if (sm->mremap)
3253                 return sm->mremap(sm, new_vma);
3254
3255         return 0;
3256 }
3257
3258 static const struct vm_operations_struct special_mapping_vmops = {
3259         .close = special_mapping_close,
3260         .fault = special_mapping_fault,
3261         .mremap = special_mapping_mremap,
3262         .name = special_mapping_name,
3263 };
3264
3265 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3266         .close = special_mapping_close,
3267         .fault = special_mapping_fault,
3268 };
3269
3270 static int special_mapping_fault(struct vm_fault *vmf)
3271 {
3272         struct vm_area_struct *vma = vmf->vma;
3273         pgoff_t pgoff;
3274         struct page **pages;
3275
3276         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3277                 pages = vma->vm_private_data;
3278         } else {
3279                 struct vm_special_mapping *sm = vma->vm_private_data;
3280
3281                 if (sm->fault)
3282                         return sm->fault(sm, vmf->vma, vmf);
3283
3284                 pages = sm->pages;
3285         }
3286
3287         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3288                 pgoff--;
3289
3290         if (*pages) {
3291                 struct page *page = *pages;
3292                 get_page(page);
3293                 vmf->page = page;
3294                 return 0;
3295         }
3296
3297         return VM_FAULT_SIGBUS;
3298 }
3299
3300 static struct vm_area_struct *__install_special_mapping(
3301         struct mm_struct *mm,
3302         unsigned long addr, unsigned long len,
3303         unsigned long vm_flags, void *priv,
3304         const struct vm_operations_struct *ops)
3305 {
3306         int ret;
3307         struct vm_area_struct *vma;
3308
3309         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3310         if (unlikely(vma == NULL))
3311                 return ERR_PTR(-ENOMEM);
3312
3313         INIT_LIST_HEAD(&vma->anon_vma_chain);
3314         vma->vm_mm = mm;
3315         vma->vm_start = addr;
3316         vma->vm_end = addr + len;
3317
3318         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3319         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3320
3321         vma->vm_ops = ops;
3322         vma->vm_private_data = priv;
3323
3324         ret = insert_vm_struct(mm, vma);
3325         if (ret)
3326                 goto out;
3327
3328         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3329
3330         perf_event_mmap(vma);
3331
3332         return vma;
3333
3334 out:
3335         kmem_cache_free(vm_area_cachep, vma);
3336         return ERR_PTR(ret);
3337 }
3338
3339 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3340         const struct vm_special_mapping *sm)
3341 {
3342         return vma->vm_private_data == sm &&
3343                 (vma->vm_ops == &special_mapping_vmops ||
3344                  vma->vm_ops == &legacy_special_mapping_vmops);
3345 }
3346
3347 /*
3348  * Called with mm->mmap_sem held for writing.
3349  * Insert a new vma covering the given region, with the given flags.
3350  * Its pages are supplied by the given array of struct page *.
3351  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3352  * The region past the last page supplied will always produce SIGBUS.
3353  * The array pointer and the pages it points to are assumed to stay alive
3354  * for as long as this mapping might exist.
3355  */
3356 struct vm_area_struct *_install_special_mapping(
3357         struct mm_struct *mm,
3358         unsigned long addr, unsigned long len,
3359         unsigned long vm_flags, const struct vm_special_mapping *spec)
3360 {
3361         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3362                                         &special_mapping_vmops);
3363 }
3364
3365 int install_special_mapping(struct mm_struct *mm,
3366                             unsigned long addr, unsigned long len,
3367                             unsigned long vm_flags, struct page **pages)
3368 {
3369         struct vm_area_struct *vma = __install_special_mapping(
3370                 mm, addr, len, vm_flags, (void *)pages,
3371                 &legacy_special_mapping_vmops);
3372
3373         return PTR_ERR_OR_ZERO(vma);
3374 }
3375
3376 static DEFINE_MUTEX(mm_all_locks_mutex);
3377
3378 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3379 {
3380         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3381                 /*
3382                  * The LSB of head.next can't change from under us
3383                  * because we hold the mm_all_locks_mutex.
3384                  */
3385                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3386                 /*
3387                  * We can safely modify head.next after taking the
3388                  * anon_vma->root->rwsem. If some other vma in this mm shares
3389                  * the same anon_vma we won't take it again.
3390                  *
3391                  * No need of atomic instructions here, head.next
3392                  * can't change from under us thanks to the
3393                  * anon_vma->root->rwsem.
3394                  */
3395                 if (__test_and_set_bit(0, (unsigned long *)
3396                                        &anon_vma->root->rb_root.rb_root.rb_node))
3397                         BUG();
3398         }
3399 }
3400
3401 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3402 {
3403         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3404                 /*
3405                  * AS_MM_ALL_LOCKS can't change from under us because
3406                  * we hold the mm_all_locks_mutex.
3407                  *
3408                  * Operations on ->flags have to be atomic because
3409                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3410                  * mm_all_locks_mutex, there may be other cpus
3411                  * changing other bitflags in parallel to us.
3412                  */
3413                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3414                         BUG();
3415                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3416         }
3417 }
3418
3419 /*
3420  * This operation locks against the VM for all pte/vma/mm related
3421  * operations that could ever happen on a certain mm. This includes
3422  * vmtruncate, try_to_unmap, and all page faults.
3423  *
3424  * The caller must take the mmap_sem in write mode before calling
3425  * mm_take_all_locks(). The caller isn't allowed to release the
3426  * mmap_sem until mm_drop_all_locks() returns.
3427  *
3428  * mmap_sem in write mode is required in order to block all operations
3429  * that could modify pagetables and free pages without need of
3430  * altering the vma layout. It's also needed in write mode to avoid new
3431  * anon_vmas to be associated with existing vmas.
3432  *
3433  * A single task can't take more than one mm_take_all_locks() in a row
3434  * or it would deadlock.
3435  *
3436  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3437  * mapping->flags avoid to take the same lock twice, if more than one
3438  * vma in this mm is backed by the same anon_vma or address_space.
3439  *
3440  * We take locks in following order, accordingly to comment at beginning
3441  * of mm/rmap.c:
3442  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3443  *     hugetlb mapping);
3444  *   - all i_mmap_rwsem locks;
3445  *   - all anon_vma->rwseml
3446  *
3447  * We can take all locks within these types randomly because the VM code
3448  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3449  * mm_all_locks_mutex.
3450  *
3451  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3452  * that may have to take thousand of locks.
3453  *
3454  * mm_take_all_locks() can fail if it's interrupted by signals.
3455  */
3456 int mm_take_all_locks(struct mm_struct *mm)
3457 {
3458         struct vm_area_struct *vma;
3459         struct anon_vma_chain *avc;
3460
3461         BUG_ON(down_read_trylock(&mm->mmap_sem));
3462
3463         mutex_lock(&mm_all_locks_mutex);
3464
3465         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3466                 i