mm: mremap: downgrade mmap_sem to read when shrinking
[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 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
104 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
105 {
106         return prot;
107 }
108 #endif
109
110 pgprot_t vm_get_page_prot(unsigned long vm_flags)
111 {
112         pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
113                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
114                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
115
116         return arch_filter_pgprot(ret);
117 }
118 EXPORT_SYMBOL(vm_get_page_prot);
119
120 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
121 {
122         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
123 }
124
125 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
126 void vma_set_page_prot(struct vm_area_struct *vma)
127 {
128         unsigned long vm_flags = vma->vm_flags;
129         pgprot_t vm_page_prot;
130
131         vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
132         if (vma_wants_writenotify(vma, vm_page_prot)) {
133                 vm_flags &= ~VM_SHARED;
134                 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
135         }
136         /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
137         WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
138 }
139
140 /*
141  * Requires inode->i_mapping->i_mmap_rwsem
142  */
143 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
144                 struct file *file, struct address_space *mapping)
145 {
146         if (vma->vm_flags & VM_DENYWRITE)
147                 atomic_inc(&file_inode(file)->i_writecount);
148         if (vma->vm_flags & VM_SHARED)
149                 mapping_unmap_writable(mapping);
150
151         flush_dcache_mmap_lock(mapping);
152         vma_interval_tree_remove(vma, &mapping->i_mmap);
153         flush_dcache_mmap_unlock(mapping);
154 }
155
156 /*
157  * Unlink a file-based vm structure from its interval tree, to hide
158  * vma from rmap and vmtruncate before freeing its page tables.
159  */
160 void unlink_file_vma(struct vm_area_struct *vma)
161 {
162         struct file *file = vma->vm_file;
163
164         if (file) {
165                 struct address_space *mapping = file->f_mapping;
166                 i_mmap_lock_write(mapping);
167                 __remove_shared_vm_struct(vma, file, mapping);
168                 i_mmap_unlock_write(mapping);
169         }
170 }
171
172 /*
173  * Close a vm structure and free it, returning the next.
174  */
175 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
176 {
177         struct vm_area_struct *next = vma->vm_next;
178
179         might_sleep();
180         if (vma->vm_ops && vma->vm_ops->close)
181                 vma->vm_ops->close(vma);
182         if (vma->vm_file)
183                 fput(vma->vm_file);
184         mpol_put(vma_policy(vma));
185         vm_area_free(vma);
186         return next;
187 }
188
189 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
190                 struct list_head *uf);
191 SYSCALL_DEFINE1(brk, unsigned long, brk)
192 {
193         unsigned long retval;
194         unsigned long newbrk, oldbrk;
195         struct mm_struct *mm = current->mm;
196         struct vm_area_struct *next;
197         unsigned long min_brk;
198         bool populate;
199         LIST_HEAD(uf);
200
201         if (down_write_killable(&mm->mmap_sem))
202                 return -EINTR;
203
204 #ifdef CONFIG_COMPAT_BRK
205         /*
206          * CONFIG_COMPAT_BRK can still be overridden by setting
207          * randomize_va_space to 2, which will still cause mm->start_brk
208          * to be arbitrarily shifted
209          */
210         if (current->brk_randomized)
211                 min_brk = mm->start_brk;
212         else
213                 min_brk = mm->end_data;
214 #else
215         min_brk = mm->start_brk;
216 #endif
217         if (brk < min_brk)
218                 goto out;
219
220         /*
221          * Check against rlimit here. If this check is done later after the test
222          * of oldbrk with newbrk then it can escape the test and let the data
223          * segment grow beyond its set limit the in case where the limit is
224          * not page aligned -Ram Gupta
225          */
226         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
227                               mm->end_data, mm->start_data))
228                 goto out;
229
230         newbrk = PAGE_ALIGN(brk);
231         oldbrk = PAGE_ALIGN(mm->brk);
232         if (oldbrk == newbrk)
233                 goto set_brk;
234
235         /* Always allow shrinking brk. */
236         if (brk <= mm->brk) {
237                 if (!do_munmap(mm, newbrk, oldbrk-newbrk, &uf))
238                         goto set_brk;
239                 goto out;
240         }
241
242         /* Check against existing mmap mappings. */
243         next = find_vma(mm, oldbrk);
244         if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
245                 goto out;
246
247         /* Ok, looks good - let it rip. */
248         if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
249                 goto out;
250
251 set_brk:
252         mm->brk = brk;
253         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
254         up_write(&mm->mmap_sem);
255         userfaultfd_unmap_complete(mm, &uf);
256         if (populate)
257                 mm_populate(oldbrk, newbrk - oldbrk);
258         return brk;
259
260 out:
261         retval = mm->brk;
262         up_write(&mm->mmap_sem);
263         return retval;
264 }
265
266 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
267 {
268         unsigned long max, prev_end, subtree_gap;
269
270         /*
271          * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
272          * allow two stack_guard_gaps between them here, and when choosing
273          * an unmapped area; whereas when expanding we only require one.
274          * That's a little inconsistent, but keeps the code here simpler.
275          */
276         max = vm_start_gap(vma);
277         if (vma->vm_prev) {
278                 prev_end = vm_end_gap(vma->vm_prev);
279                 if (max > prev_end)
280                         max -= prev_end;
281                 else
282                         max = 0;
283         }
284         if (vma->vm_rb.rb_left) {
285                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
286                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
287                 if (subtree_gap > max)
288                         max = subtree_gap;
289         }
290         if (vma->vm_rb.rb_right) {
291                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
292                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
293                 if (subtree_gap > max)
294                         max = subtree_gap;
295         }
296         return max;
297 }
298
299 #ifdef CONFIG_DEBUG_VM_RB
300 static int browse_rb(struct mm_struct *mm)
301 {
302         struct rb_root *root = &mm->mm_rb;
303         int i = 0, j, bug = 0;
304         struct rb_node *nd, *pn = NULL;
305         unsigned long prev = 0, pend = 0;
306
307         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308                 struct vm_area_struct *vma;
309                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310                 if (vma->vm_start < prev) {
311                         pr_emerg("vm_start %lx < prev %lx\n",
312                                   vma->vm_start, prev);
313                         bug = 1;
314                 }
315                 if (vma->vm_start < pend) {
316                         pr_emerg("vm_start %lx < pend %lx\n",
317                                   vma->vm_start, pend);
318                         bug = 1;
319                 }
320                 if (vma->vm_start > vma->vm_end) {
321                         pr_emerg("vm_start %lx > vm_end %lx\n",
322                                   vma->vm_start, vma->vm_end);
323                         bug = 1;
324                 }
325                 spin_lock(&mm->page_table_lock);
326                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
327                         pr_emerg("free gap %lx, correct %lx\n",
328                                vma->rb_subtree_gap,
329                                vma_compute_subtree_gap(vma));
330                         bug = 1;
331                 }
332                 spin_unlock(&mm->page_table_lock);
333                 i++;
334                 pn = nd;
335                 prev = vma->vm_start;
336                 pend = vma->vm_end;
337         }
338         j = 0;
339         for (nd = pn; nd; nd = rb_prev(nd))
340                 j++;
341         if (i != j) {
342                 pr_emerg("backwards %d, forwards %d\n", j, i);
343                 bug = 1;
344         }
345         return bug ? -1 : i;
346 }
347
348 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
349 {
350         struct rb_node *nd;
351
352         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
353                 struct vm_area_struct *vma;
354                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
355                 VM_BUG_ON_VMA(vma != ignore &&
356                         vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
357                         vma);
358         }
359 }
360
361 static void validate_mm(struct mm_struct *mm)
362 {
363         int bug = 0;
364         int i = 0;
365         unsigned long highest_address = 0;
366         struct vm_area_struct *vma = mm->mmap;
367
368         while (vma) {
369                 struct anon_vma *anon_vma = vma->anon_vma;
370                 struct anon_vma_chain *avc;
371
372                 if (anon_vma) {
373                         anon_vma_lock_read(anon_vma);
374                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
375                                 anon_vma_interval_tree_verify(avc);
376                         anon_vma_unlock_read(anon_vma);
377                 }
378
379                 highest_address = vm_end_gap(vma);
380                 vma = vma->vm_next;
381                 i++;
382         }
383         if (i != mm->map_count) {
384                 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
385                 bug = 1;
386         }
387         if (highest_address != mm->highest_vm_end) {
388                 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
389                           mm->highest_vm_end, highest_address);
390                 bug = 1;
391         }
392         i = browse_rb(mm);
393         if (i != mm->map_count) {
394                 if (i != -1)
395                         pr_emerg("map_count %d rb %d\n", mm->map_count, i);
396                 bug = 1;
397         }
398         VM_BUG_ON_MM(bug, mm);
399 }
400 #else
401 #define validate_mm_rb(root, ignore) do { } while (0)
402 #define validate_mm(mm) do { } while (0)
403 #endif
404
405 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
406                      unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
407
408 /*
409  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
410  * vma->vm_prev->vm_end values changed, without modifying the vma's position
411  * in the rbtree.
412  */
413 static void vma_gap_update(struct vm_area_struct *vma)
414 {
415         /*
416          * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
417          * function that does exacltly what we want.
418          */
419         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
420 }
421
422 static inline void vma_rb_insert(struct vm_area_struct *vma,
423                                  struct rb_root *root)
424 {
425         /* All rb_subtree_gap values must be consistent prior to insertion */
426         validate_mm_rb(root, NULL);
427
428         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
429 }
430
431 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
432 {
433         /*
434          * Note rb_erase_augmented is a fairly large inline function,
435          * so make sure we instantiate it only once with our desired
436          * augmented rbtree callbacks.
437          */
438         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
439 }
440
441 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
442                                                 struct rb_root *root,
443                                                 struct vm_area_struct *ignore)
444 {
445         /*
446          * All rb_subtree_gap values must be consistent prior to erase,
447          * with the possible exception of the "next" vma being erased if
448          * next->vm_start was reduced.
449          */
450         validate_mm_rb(root, ignore);
451
452         __vma_rb_erase(vma, root);
453 }
454
455 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
456                                          struct rb_root *root)
457 {
458         /*
459          * All rb_subtree_gap values must be consistent prior to erase,
460          * with the possible exception of the vma being erased.
461          */
462         validate_mm_rb(root, vma);
463
464         __vma_rb_erase(vma, root);
465 }
466
467 /*
468  * vma has some anon_vma assigned, and is already inserted on that
469  * anon_vma's interval trees.
470  *
471  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
472  * vma must be removed from the anon_vma's interval trees using
473  * anon_vma_interval_tree_pre_update_vma().
474  *
475  * After the update, the vma will be reinserted using
476  * anon_vma_interval_tree_post_update_vma().
477  *
478  * The entire update must be protected by exclusive mmap_sem and by
479  * the root anon_vma's mutex.
480  */
481 static inline void
482 anon_vma_interval_tree_pre_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_remove(avc, &avc->anon_vma->rb_root);
488 }
489
490 static inline void
491 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
492 {
493         struct anon_vma_chain *avc;
494
495         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
496                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
497 }
498
499 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
500                 unsigned long end, struct vm_area_struct **pprev,
501                 struct rb_node ***rb_link, struct rb_node **rb_parent)
502 {
503         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
504
505         __rb_link = &mm->mm_rb.rb_node;
506         rb_prev = __rb_parent = NULL;
507
508         while (*__rb_link) {
509                 struct vm_area_struct *vma_tmp;
510
511                 __rb_parent = *__rb_link;
512                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
513
514                 if (vma_tmp->vm_end > addr) {
515                         /* Fail if an existing vma overlaps the area */
516                         if (vma_tmp->vm_start < end)
517                                 return -ENOMEM;
518                         __rb_link = &__rb_parent->rb_left;
519                 } else {
520                         rb_prev = __rb_parent;
521                         __rb_link = &__rb_parent->rb_right;
522                 }
523         }
524
525         *pprev = NULL;
526         if (rb_prev)
527                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
528         *rb_link = __rb_link;
529         *rb_parent = __rb_parent;
530         return 0;
531 }
532
533 static unsigned long count_vma_pages_range(struct mm_struct *mm,
534                 unsigned long addr, unsigned long end)
535 {
536         unsigned long nr_pages = 0;
537         struct vm_area_struct *vma;
538
539         /* Find first overlaping mapping */
540         vma = find_vma_intersection(mm, addr, end);
541         if (!vma)
542                 return 0;
543
544         nr_pages = (min(end, vma->vm_end) -
545                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
546
547         /* Iterate over the rest of the overlaps */
548         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
549                 unsigned long overlap_len;
550
551                 if (vma->vm_start > end)
552                         break;
553
554                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
555                 nr_pages += overlap_len >> PAGE_SHIFT;
556         }
557
558         return nr_pages;
559 }
560
561 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
562                 struct rb_node **rb_link, struct rb_node *rb_parent)
563 {
564         /* Update tracking information for the gap following the new vma. */
565         if (vma->vm_next)
566                 vma_gap_update(vma->vm_next);
567         else
568                 mm->highest_vm_end = vm_end_gap(vma);
569
570         /*
571          * vma->vm_prev wasn't known when we followed the rbtree to find the
572          * correct insertion point for that vma. As a result, we could not
573          * update the vma vm_rb parents rb_subtree_gap values on the way down.
574          * So, we first insert the vma with a zero rb_subtree_gap value
575          * (to be consistent with what we did on the way down), and then
576          * immediately update the gap to the correct value. Finally we
577          * rebalance the rbtree after all augmented values have been set.
578          */
579         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
580         vma->rb_subtree_gap = 0;
581         vma_gap_update(vma);
582         vma_rb_insert(vma, &mm->mm_rb);
583 }
584
585 static void __vma_link_file(struct vm_area_struct *vma)
586 {
587         struct file *file;
588
589         file = vma->vm_file;
590         if (file) {
591                 struct address_space *mapping = file->f_mapping;
592
593                 if (vma->vm_flags & VM_DENYWRITE)
594                         atomic_dec(&file_inode(file)->i_writecount);
595                 if (vma->vm_flags & VM_SHARED)
596                         atomic_inc(&mapping->i_mmap_writable);
597
598                 flush_dcache_mmap_lock(mapping);
599                 vma_interval_tree_insert(vma, &mapping->i_mmap);
600                 flush_dcache_mmap_unlock(mapping);
601         }
602 }
603
604 static void
605 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
606         struct vm_area_struct *prev, struct rb_node **rb_link,
607         struct rb_node *rb_parent)
608 {
609         __vma_link_list(mm, vma, prev, rb_parent);
610         __vma_link_rb(mm, vma, rb_link, rb_parent);
611 }
612
613 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
614                         struct vm_area_struct *prev, struct rb_node **rb_link,
615                         struct rb_node *rb_parent)
616 {
617         struct address_space *mapping = NULL;
618
619         if (vma->vm_file) {
620                 mapping = vma->vm_file->f_mapping;
621                 i_mmap_lock_write(mapping);
622         }
623
624         __vma_link(mm, vma, prev, rb_link, rb_parent);
625         __vma_link_file(vma);
626
627         if (mapping)
628                 i_mmap_unlock_write(mapping);
629
630         mm->map_count++;
631         validate_mm(mm);
632 }
633
634 /*
635  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
636  * mm's list and rbtree.  It has already been inserted into the interval tree.
637  */
638 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
639 {
640         struct vm_area_struct *prev;
641         struct rb_node **rb_link, *rb_parent;
642
643         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
644                            &prev, &rb_link, &rb_parent))
645                 BUG();
646         __vma_link(mm, vma, prev, rb_link, rb_parent);
647         mm->map_count++;
648 }
649
650 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
651                                                 struct vm_area_struct *vma,
652                                                 struct vm_area_struct *prev,
653                                                 bool has_prev,
654                                                 struct vm_area_struct *ignore)
655 {
656         struct vm_area_struct *next;
657
658         vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
659         next = vma->vm_next;
660         if (has_prev)
661                 prev->vm_next = next;
662         else {
663                 prev = vma->vm_prev;
664                 if (prev)
665                         prev->vm_next = next;
666                 else
667                         mm->mmap = next;
668         }
669         if (next)
670                 next->vm_prev = prev;
671
672         /* Kill the cache */
673         vmacache_invalidate(mm);
674 }
675
676 static inline void __vma_unlink_prev(struct mm_struct *mm,
677                                      struct vm_area_struct *vma,
678                                      struct vm_area_struct *prev)
679 {
680         __vma_unlink_common(mm, vma, prev, true, vma);
681 }
682
683 /*
684  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
685  * is already present in an i_mmap tree without adjusting the tree.
686  * The following helper function should be used when such adjustments
687  * are necessary.  The "insert" vma (if any) is to be inserted
688  * before we drop the necessary locks.
689  */
690 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
691         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
692         struct vm_area_struct *expand)
693 {
694         struct mm_struct *mm = vma->vm_mm;
695         struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
696         struct address_space *mapping = NULL;
697         struct rb_root_cached *root = NULL;
698         struct anon_vma *anon_vma = NULL;
699         struct file *file = vma->vm_file;
700         bool start_changed = false, end_changed = false;
701         long adjust_next = 0;
702         int remove_next = 0;
703
704         if (next && !insert) {
705                 struct vm_area_struct *exporter = NULL, *importer = NULL;
706
707                 if (end >= next->vm_end) {
708                         /*
709                          * vma expands, overlapping all the next, and
710                          * perhaps the one after too (mprotect case 6).
711                          * The only other cases that gets here are
712                          * case 1, case 7 and case 8.
713                          */
714                         if (next == expand) {
715                                 /*
716                                  * The only case where we don't expand "vma"
717                                  * and we expand "next" instead is case 8.
718                                  */
719                                 VM_WARN_ON(end != next->vm_end);
720                                 /*
721                                  * remove_next == 3 means we're
722                                  * removing "vma" and that to do so we
723                                  * swapped "vma" and "next".
724                                  */
725                                 remove_next = 3;
726                                 VM_WARN_ON(file != next->vm_file);
727                                 swap(vma, next);
728                         } else {
729                                 VM_WARN_ON(expand != vma);
730                                 /*
731                                  * case 1, 6, 7, remove_next == 2 is case 6,
732                                  * remove_next == 1 is case 1 or 7.
733                                  */
734                                 remove_next = 1 + (end > next->vm_end);
735                                 VM_WARN_ON(remove_next == 2 &&
736                                            end != next->vm_next->vm_end);
737                                 VM_WARN_ON(remove_next == 1 &&
738                                            end != next->vm_end);
739                                 /* trim end to next, for case 6 first pass */
740                                 end = next->vm_end;
741                         }
742
743                         exporter = next;
744                         importer = vma;
745
746                         /*
747                          * If next doesn't have anon_vma, import from vma after
748                          * next, if the vma overlaps with it.
749                          */
750                         if (remove_next == 2 && !next->anon_vma)
751                                 exporter = next->vm_next;
752
753                 } else if (end > next->vm_start) {
754                         /*
755                          * vma expands, overlapping part of the next:
756                          * mprotect case 5 shifting the boundary up.
757                          */
758                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
759                         exporter = next;
760                         importer = vma;
761                         VM_WARN_ON(expand != importer);
762                 } else if (end < vma->vm_end) {
763                         /*
764                          * vma shrinks, and !insert tells it's not
765                          * split_vma inserting another: so it must be
766                          * mprotect case 4 shifting the boundary down.
767                          */
768                         adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
769                         exporter = vma;
770                         importer = next;
771                         VM_WARN_ON(expand != importer);
772                 }
773
774                 /*
775                  * Easily overlooked: when mprotect shifts the boundary,
776                  * make sure the expanding vma has anon_vma set if the
777                  * shrinking vma had, to cover any anon pages imported.
778                  */
779                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
780                         int error;
781
782                         importer->anon_vma = exporter->anon_vma;
783                         error = anon_vma_clone(importer, exporter);
784                         if (error)
785                                 return error;
786                 }
787         }
788 again:
789         vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
790
791         if (file) {
792                 mapping = file->f_mapping;
793                 root = &mapping->i_mmap;
794                 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
795
796                 if (adjust_next)
797                         uprobe_munmap(next, next->vm_start, next->vm_end);
798
799                 i_mmap_lock_write(mapping);
800                 if (insert) {
801                         /*
802                          * Put into interval tree now, so instantiated pages
803                          * are visible to arm/parisc __flush_dcache_page
804                          * throughout; but we cannot insert into address
805                          * space until vma start or end is updated.
806                          */
807                         __vma_link_file(insert);
808                 }
809         }
810
811         anon_vma = vma->anon_vma;
812         if (!anon_vma && adjust_next)
813                 anon_vma = next->anon_vma;
814         if (anon_vma) {
815                 VM_WARN_ON(adjust_next && next->anon_vma &&
816                            anon_vma != next->anon_vma);
817                 anon_vma_lock_write(anon_vma);
818                 anon_vma_interval_tree_pre_update_vma(vma);
819                 if (adjust_next)
820                         anon_vma_interval_tree_pre_update_vma(next);
821         }
822
823         if (root) {
824                 flush_dcache_mmap_lock(mapping);
825                 vma_interval_tree_remove(vma, root);
826                 if (adjust_next)
827                         vma_interval_tree_remove(next, root);
828         }
829
830         if (start != vma->vm_start) {
831                 vma->vm_start = start;
832                 start_changed = true;
833         }
834         if (end != vma->vm_end) {
835                 vma->vm_end = end;
836                 end_changed = true;
837         }
838         vma->vm_pgoff = pgoff;
839         if (adjust_next) {
840                 next->vm_start += adjust_next << PAGE_SHIFT;
841                 next->vm_pgoff += adjust_next;
842         }
843
844         if (root) {
845                 if (adjust_next)
846                         vma_interval_tree_insert(next, root);
847                 vma_interval_tree_insert(vma, root);
848                 flush_dcache_mmap_unlock(mapping);
849         }
850
851         if (remove_next) {
852                 /*
853                  * vma_merge has merged next into vma, and needs
854                  * us to remove next before dropping the locks.
855                  */
856                 if (remove_next != 3)
857                         __vma_unlink_prev(mm, next, vma);
858                 else
859                         /*
860                          * vma is not before next if they've been
861                          * swapped.
862                          *
863                          * pre-swap() next->vm_start was reduced so
864                          * tell validate_mm_rb to ignore pre-swap()
865                          * "next" (which is stored in post-swap()
866                          * "vma").
867                          */
868                         __vma_unlink_common(mm, next, NULL, false, vma);
869                 if (file)
870                         __remove_shared_vm_struct(next, file, mapping);
871         } else if (insert) {
872                 /*
873                  * split_vma has split insert from vma, and needs
874                  * us to insert it before dropping the locks
875                  * (it may either follow vma or precede it).
876                  */
877                 __insert_vm_struct(mm, insert);
878         } else {
879                 if (start_changed)
880                         vma_gap_update(vma);
881                 if (end_changed) {
882                         if (!next)
883                                 mm->highest_vm_end = vm_end_gap(vma);
884                         else if (!adjust_next)
885                                 vma_gap_update(next);
886                 }
887         }
888
889         if (anon_vma) {
890                 anon_vma_interval_tree_post_update_vma(vma);
891                 if (adjust_next)
892                         anon_vma_interval_tree_post_update_vma(next);
893                 anon_vma_unlock_write(anon_vma);
894         }
895         if (mapping)
896                 i_mmap_unlock_write(mapping);
897
898         if (root) {
899                 uprobe_mmap(vma);
900
901                 if (adjust_next)
902                         uprobe_mmap(next);
903         }
904
905         if (remove_next) {
906                 if (file) {
907                         uprobe_munmap(next, next->vm_start, next->vm_end);
908                         fput(file);
909                 }
910                 if (next->anon_vma)
911                         anon_vma_merge(vma, next);
912                 mm->map_count--;
913                 mpol_put(vma_policy(next));
914                 vm_area_free(next);
915                 /*
916                  * In mprotect's case 6 (see comments on vma_merge),
917                  * we must remove another next too. It would clutter
918                  * up the code too much to do both in one go.
919                  */
920                 if (remove_next != 3) {
921                         /*
922                          * If "next" was removed and vma->vm_end was
923                          * expanded (up) over it, in turn
924                          * "next->vm_prev->vm_end" changed and the
925                          * "vma->vm_next" gap must be updated.
926                          */
927                         next = vma->vm_next;
928                 } else {
929                         /*
930                          * For the scope of the comment "next" and
931                          * "vma" considered pre-swap(): if "vma" was
932                          * removed, next->vm_start was expanded (down)
933                          * over it and the "next" gap must be updated.
934                          * Because of the swap() the post-swap() "vma"
935                          * actually points to pre-swap() "next"
936                          * (post-swap() "next" as opposed is now a
937                          * dangling pointer).
938                          */
939                         next = vma;
940                 }
941                 if (remove_next == 2) {
942                         remove_next = 1;
943                         end = next->vm_end;
944                         goto again;
945                 }
946                 else if (next)
947                         vma_gap_update(next);
948                 else {
949                         /*
950                          * If remove_next == 2 we obviously can't
951                          * reach this path.
952                          *
953                          * If remove_next == 3 we can't reach this
954                          * path because pre-swap() next is always not
955                          * NULL. pre-swap() "next" is not being
956                          * removed and its next->vm_end is not altered
957                          * (and furthermore "end" already matches
958                          * next->vm_end in remove_next == 3).
959                          *
960                          * We reach this only in the remove_next == 1
961                          * case if the "next" vma that was removed was
962                          * the highest vma of the mm. However in such
963                          * case next->vm_end == "end" and the extended
964                          * "vma" has vma->vm_end == next->vm_end so
965                          * mm->highest_vm_end doesn't need any update
966                          * in remove_next == 1 case.
967                          */
968                         VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
969                 }
970         }
971         if (insert && file)
972                 uprobe_mmap(insert);
973
974         validate_mm(mm);
975
976         return 0;
977 }
978
979 /*
980  * If the vma has a ->close operation then the driver probably needs to release
981  * per-vma resources, so we don't attempt to merge those.
982  */
983 static inline int is_mergeable_vma(struct vm_area_struct *vma,
984                                 struct file *file, unsigned long vm_flags,
985                                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
986 {
987         /*
988          * VM_SOFTDIRTY should not prevent from VMA merging, if we
989          * match the flags but dirty bit -- the caller should mark
990          * merged VMA as dirty. If dirty bit won't be excluded from
991          * comparison, we increase pressue on the memory system forcing
992          * the kernel to generate new VMAs when old one could be
993          * extended instead.
994          */
995         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
996                 return 0;
997         if (vma->vm_file != file)
998                 return 0;
999         if (vma->vm_ops && vma->vm_ops->close)
1000                 return 0;
1001         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1002                 return 0;
1003         return 1;
1004 }
1005
1006 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1007                                         struct anon_vma *anon_vma2,
1008                                         struct vm_area_struct *vma)
1009 {
1010         /*
1011          * The list_is_singular() test is to avoid merging VMA cloned from
1012          * parents. This can improve scalability caused by anon_vma lock.
1013          */
1014         if ((!anon_vma1 || !anon_vma2) && (!vma ||
1015                 list_is_singular(&vma->anon_vma_chain)))
1016                 return 1;
1017         return anon_vma1 == anon_vma2;
1018 }
1019
1020 /*
1021  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1022  * in front of (at a lower virtual address and file offset than) the vma.
1023  *
1024  * We cannot merge two vmas if they have differently assigned (non-NULL)
1025  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1026  *
1027  * We don't check here for the merged mmap wrapping around the end of pagecache
1028  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1029  * wrap, nor mmaps which cover the final page at index -1UL.
1030  */
1031 static int
1032 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1033                      struct anon_vma *anon_vma, struct file *file,
1034                      pgoff_t vm_pgoff,
1035                      struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1036 {
1037         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1038             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1039                 if (vma->vm_pgoff == vm_pgoff)
1040                         return 1;
1041         }
1042         return 0;
1043 }
1044
1045 /*
1046  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1047  * beyond (at a higher virtual address and file offset than) the vma.
1048  *
1049  * We cannot merge two vmas if they have differently assigned (non-NULL)
1050  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1051  */
1052 static int
1053 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1054                     struct anon_vma *anon_vma, struct file *file,
1055                     pgoff_t vm_pgoff,
1056                     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1057 {
1058         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1059             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1060                 pgoff_t vm_pglen;
1061                 vm_pglen = vma_pages(vma);
1062                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1063                         return 1;
1064         }
1065         return 0;
1066 }
1067
1068 /*
1069  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1070  * whether that can be merged with its predecessor or its successor.
1071  * Or both (it neatly fills a hole).
1072  *
1073  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1074  * certain not to be mapped by the time vma_merge is called; but when
1075  * called for mprotect, it is certain to be already mapped (either at
1076  * an offset within prev, or at the start of next), and the flags of
1077  * this area are about to be changed to vm_flags - and the no-change
1078  * case has already been eliminated.
1079  *
1080  * The following mprotect cases have to be considered, where AAAA is
1081  * the area passed down from mprotect_fixup, never extending beyond one
1082  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1083  *
1084  *     AAAA             AAAA                AAAA          AAAA
1085  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
1086  *    cannot merge    might become    might become    might become
1087  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
1088  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
1089  *    mremap move:                                    PPPPXXXXXXXX 8
1090  *        AAAA
1091  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
1092  *    might become    case 1 below    case 2 below    case 3 below
1093  *
1094  * It is important for case 8 that the the vma NNNN overlapping the
1095  * region AAAA is never going to extended over XXXX. Instead XXXX must
1096  * be extended in region AAAA and NNNN must be removed. This way in
1097  * all cases where vma_merge succeeds, the moment vma_adjust drops the
1098  * rmap_locks, the properties of the merged vma will be already
1099  * correct for the whole merged range. Some of those properties like
1100  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1101  * be correct for the whole merged range immediately after the
1102  * rmap_locks are released. Otherwise if XXXX would be removed and
1103  * NNNN would be extended over the XXXX range, remove_migration_ptes
1104  * or other rmap walkers (if working on addresses beyond the "end"
1105  * parameter) may establish ptes with the wrong permissions of NNNN
1106  * instead of the right permissions of XXXX.
1107  */
1108 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1109                         struct vm_area_struct *prev, unsigned long addr,
1110                         unsigned long end, unsigned long vm_flags,
1111                         struct anon_vma *anon_vma, struct file *file,
1112                         pgoff_t pgoff, struct mempolicy *policy,
1113                         struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1114 {
1115         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1116         struct vm_area_struct *area, *next;
1117         int err;
1118
1119         /*
1120          * We later require that vma->vm_flags == vm_flags,
1121          * so this tests vma->vm_flags & VM_SPECIAL, too.
1122          */
1123         if (vm_flags & VM_SPECIAL)
1124                 return NULL;
1125
1126         if (prev)
1127                 next = prev->vm_next;
1128         else
1129                 next = mm->mmap;
1130         area = next;
1131         if (area && area->vm_end == end)                /* cases 6, 7, 8 */
1132                 next = next->vm_next;
1133
1134         /* verify some invariant that must be enforced by the caller */
1135         VM_WARN_ON(prev && addr <= prev->vm_start);
1136         VM_WARN_ON(area && end > area->vm_end);
1137         VM_WARN_ON(addr >= end);
1138
1139         /*
1140          * Can it merge with the predecessor?
1141          */
1142         if (prev && prev->vm_end == addr &&
1143                         mpol_equal(vma_policy(prev), policy) &&
1144                         can_vma_merge_after(prev, vm_flags,
1145                                             anon_vma, file, pgoff,
1146                                             vm_userfaultfd_ctx)) {
1147                 /*
1148                  * OK, it can.  Can we now merge in the successor as well?
1149                  */
1150                 if (next && end == next->vm_start &&
1151                                 mpol_equal(policy, vma_policy(next)) &&
1152                                 can_vma_merge_before(next, vm_flags,
1153                                                      anon_vma, file,
1154                                                      pgoff+pglen,
1155                                                      vm_userfaultfd_ctx) &&
1156                                 is_mergeable_anon_vma(prev->anon_vma,
1157                                                       next->anon_vma, NULL)) {
1158                                                         /* cases 1, 6 */
1159                         err = __vma_adjust(prev, prev->vm_start,
1160                                          next->vm_end, prev->vm_pgoff, NULL,
1161                                          prev);
1162                 } else                                  /* cases 2, 5, 7 */
1163                         err = __vma_adjust(prev, prev->vm_start,
1164                                          end, prev->vm_pgoff, NULL, prev);
1165                 if (err)
1166                         return NULL;
1167                 khugepaged_enter_vma_merge(prev, vm_flags);
1168                 return prev;
1169         }
1170
1171         /*
1172          * Can this new request be merged in front of next?
1173          */
1174         if (next && end == next->vm_start &&
1175                         mpol_equal(policy, vma_policy(next)) &&
1176                         can_vma_merge_before(next, vm_flags,
1177                                              anon_vma, file, pgoff+pglen,
1178                                              vm_userfaultfd_ctx)) {
1179                 if (prev && addr < prev->vm_end)        /* case 4 */
1180                         err = __vma_adjust(prev, prev->vm_start,
1181                                          addr, prev->vm_pgoff, NULL, next);
1182                 else {                                  /* cases 3, 8 */
1183                         err = __vma_adjust(area, addr, next->vm_end,
1184                                          next->vm_pgoff - pglen, NULL, next);
1185                         /*
1186                          * In case 3 area is already equal to next and
1187                          * this is a noop, but in case 8 "area" has
1188                          * been removed and next was expanded over it.
1189                          */
1190                         area = next;
1191                 }
1192                 if (err)
1193                         return NULL;
1194                 khugepaged_enter_vma_merge(area, vm_flags);
1195                 return area;
1196         }
1197
1198         return NULL;
1199 }
1200
1201 /*
1202  * Rough compatbility check to quickly see if it's even worth looking
1203  * at sharing an anon_vma.
1204  *
1205  * They need to have the same vm_file, and the flags can only differ
1206  * in things that mprotect may change.
1207  *
1208  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1209  * we can merge the two vma's. For example, we refuse to merge a vma if
1210  * there is a vm_ops->close() function, because that indicates that the
1211  * driver is doing some kind of reference counting. But that doesn't
1212  * really matter for the anon_vma sharing case.
1213  */
1214 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1215 {
1216         return a->vm_end == b->vm_start &&
1217                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1218                 a->vm_file == b->vm_file &&
1219                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1220                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1221 }
1222
1223 /*
1224  * Do some basic sanity checking to see if we can re-use the anon_vma
1225  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1226  * the same as 'old', the other will be the new one that is trying
1227  * to share the anon_vma.
1228  *
1229  * NOTE! This runs with mm_sem held for reading, so it is possible that
1230  * the anon_vma of 'old' is concurrently in the process of being set up
1231  * by another page fault trying to merge _that_. But that's ok: if it
1232  * is being set up, that automatically means that it will be a singleton
1233  * acceptable for merging, so we can do all of this optimistically. But
1234  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1235  *
1236  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1237  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1238  * is to return an anon_vma that is "complex" due to having gone through
1239  * a fork).
1240  *
1241  * We also make sure that the two vma's are compatible (adjacent,
1242  * and with the same memory policies). That's all stable, even with just
1243  * a read lock on the mm_sem.
1244  */
1245 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1246 {
1247         if (anon_vma_compatible(a, b)) {
1248                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1249
1250                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1251                         return anon_vma;
1252         }
1253         return NULL;
1254 }
1255
1256 /*
1257  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1258  * neighbouring vmas for a suitable anon_vma, before it goes off
1259  * to allocate a new anon_vma.  It checks because a repetitive
1260  * sequence of mprotects and faults may otherwise lead to distinct
1261  * anon_vmas being allocated, preventing vma merge in subsequent
1262  * mprotect.
1263  */
1264 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1265 {
1266         struct anon_vma *anon_vma;
1267         struct vm_area_struct *near;
1268
1269         near = vma->vm_next;
1270         if (!near)
1271                 goto try_prev;
1272
1273         anon_vma = reusable_anon_vma(near, vma, near);
1274         if (anon_vma)
1275                 return anon_vma;
1276 try_prev:
1277         near = vma->vm_prev;
1278         if (!near)
1279                 goto none;
1280
1281         anon_vma = reusable_anon_vma(near, near, vma);
1282         if (anon_vma)
1283                 return anon_vma;
1284 none:
1285         /*
1286          * There's no absolute need to look only at touching neighbours:
1287          * we could search further afield for "compatible" anon_vmas.
1288          * But it would probably just be a waste of time searching,
1289          * or lead to too many vmas hanging off the same anon_vma.
1290          * We're trying to allow mprotect remerging later on,
1291          * not trying to minimize memory used for anon_vmas.
1292          */
1293         return NULL;
1294 }
1295
1296 /*
1297  * If a hint addr is less than mmap_min_addr change hint to be as
1298  * low as possible but still greater than mmap_min_addr
1299  */
1300 static inline unsigned long round_hint_to_min(unsigned long hint)
1301 {
1302         hint &= PAGE_MASK;
1303         if (((void *)hint != NULL) &&
1304             (hint < mmap_min_addr))
1305                 return PAGE_ALIGN(mmap_min_addr);
1306         return hint;
1307 }
1308
1309 static inline int mlock_future_check(struct mm_struct *mm,
1310                                      unsigned long flags,
1311                                      unsigned long len)
1312 {
1313         unsigned long locked, lock_limit;
1314
1315         /*  mlock MCL_FUTURE? */
1316         if (flags & VM_LOCKED) {
1317                 locked = len >> PAGE_SHIFT;
1318                 locked += mm->locked_vm;
1319                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1320                 lock_limit >>= PAGE_SHIFT;
1321                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1322                         return -EAGAIN;
1323         }
1324         return 0;
1325 }
1326
1327 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1328 {
1329         if (S_ISREG(inode->i_mode))
1330                 return MAX_LFS_FILESIZE;
1331
1332         if (S_ISBLK(inode->i_mode))
1333                 return MAX_LFS_FILESIZE;
1334
1335         /* Special "we do even unsigned file positions" case */
1336         if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1337                 return 0;
1338
1339         /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1340         return ULONG_MAX;
1341 }
1342
1343 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1344                                 unsigned long pgoff, unsigned long len)
1345 {
1346         u64 maxsize = file_mmap_size_max(file, inode);
1347
1348         if (maxsize && len > maxsize)
1349                 return false;
1350         maxsize -= len;
1351         if (pgoff > maxsize >> PAGE_SHIFT)
1352                 return false;
1353         return true;
1354 }
1355
1356 /*
1357  * The caller must hold down_write(&current->mm->mmap_sem).
1358  */
1359 unsigned long do_mmap(struct file *file, unsigned long addr,
1360                         unsigned long len, unsigned long prot,
1361                         unsigned long flags, vm_flags_t vm_flags,
1362                         unsigned long pgoff, unsigned long *populate,
1363                         struct list_head *uf)
1364 {
1365         struct mm_struct *mm = current->mm;
1366         int pkey = 0;
1367
1368         *populate = 0;
1369
1370         if (!len)
1371                 return -EINVAL;
1372
1373         /*
1374          * Does the application expect PROT_READ to imply PROT_EXEC?
1375          *
1376          * (the exception is when the underlying filesystem is noexec
1377          *  mounted, in which case we dont add PROT_EXEC.)
1378          */
1379         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1380                 if (!(file && path_noexec(&file->f_path)))
1381                         prot |= PROT_EXEC;
1382
1383         /* force arch specific MAP_FIXED handling in get_unmapped_area */
1384         if (flags & MAP_FIXED_NOREPLACE)
1385                 flags |= MAP_FIXED;
1386
1387         if (!(flags & MAP_FIXED))
1388                 addr = round_hint_to_min(addr);
1389
1390         /* Careful about overflows.. */
1391         len = PAGE_ALIGN(len);
1392         if (!len)
1393                 return -ENOMEM;
1394
1395         /* offset overflow? */
1396         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1397                 return -EOVERFLOW;
1398
1399         /* Too many mappings? */
1400         if (mm->map_count > sysctl_max_map_count)
1401                 return -ENOMEM;
1402
1403         /* Obtain the address to map to. we verify (or select) it and ensure
1404          * that it represents a valid section of the address space.
1405          */
1406         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1407         if (offset_in_page(addr))
1408                 return addr;
1409
1410         if (flags & MAP_FIXED_NOREPLACE) {
1411                 struct vm_area_struct *vma = find_vma(mm, addr);
1412
1413                 if (vma && vma->vm_start < addr + len)
1414                         return -EEXIST;
1415         }
1416
1417         if (prot == PROT_EXEC) {
1418                 pkey = execute_only_pkey(mm);
1419                 if (pkey < 0)
1420                         pkey = 0;
1421         }
1422
1423         /* Do simple checking here so the lower-level routines won't have
1424          * to. we assume access permissions have been handled by the open
1425          * of the memory object, so we don't do any here.
1426          */
1427         vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1428                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1429
1430         if (flags & MAP_LOCKED)
1431                 if (!can_do_mlock())
1432                         return -EPERM;
1433
1434         if (mlock_future_check(mm, vm_flags, len))
1435                 return -EAGAIN;
1436
1437         if (file) {
1438                 struct inode *inode = file_inode(file);
1439                 unsigned long flags_mask;
1440
1441                 if (!file_mmap_ok(file, inode, pgoff, len))
1442                         return -EOVERFLOW;
1443
1444                 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1445
1446                 switch (flags & MAP_TYPE) {
1447                 case MAP_SHARED:
1448                         /*
1449                          * Force use of MAP_SHARED_VALIDATE with non-legacy
1450                          * flags. E.g. MAP_SYNC is dangerous to use with
1451                          * MAP_SHARED as you don't know which consistency model
1452                          * you will get. We silently ignore unsupported flags
1453                          * with MAP_SHARED to preserve backward compatibility.
1454                          */
1455                         flags &= LEGACY_MAP_MASK;
1456                         /* fall through */
1457                 case MAP_SHARED_VALIDATE:
1458                         if (flags & ~flags_mask)
1459                                 return -EOPNOTSUPP;
1460                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1461                                 return -EACCES;
1462
1463                         /*
1464                          * Make sure we don't allow writing to an append-only
1465                          * file..
1466                          */
1467                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1468                                 return -EACCES;
1469
1470                         /*
1471                          * Make sure there are no mandatory locks on the file.
1472                          */
1473                         if (locks_verify_locked(file))
1474                                 return -EAGAIN;
1475
1476                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1477                         if (!(file->f_mode & FMODE_WRITE))
1478                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1479
1480                         /* fall through */
1481                 case MAP_PRIVATE:
1482                         if (!(file->f_mode & FMODE_READ))
1483                                 return -EACCES;
1484                         if (path_noexec(&file->f_path)) {
1485                                 if (vm_flags & VM_EXEC)
1486                                         return -EPERM;
1487                                 vm_flags &= ~VM_MAYEXEC;
1488                         }
1489
1490                         if (!file->f_op->mmap)
1491                                 return -ENODEV;
1492                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1493                                 return -EINVAL;
1494                         break;
1495
1496                 default:
1497                         return -EINVAL;
1498                 }
1499         } else {
1500                 switch (flags & MAP_TYPE) {
1501                 case MAP_SHARED:
1502                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1503                                 return -EINVAL;
1504                         /*
1505                          * Ignore pgoff.
1506                          */
1507                         pgoff = 0;
1508                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1509                         break;
1510                 case MAP_PRIVATE:
1511                         /*
1512                          * Set pgoff according to addr for anon_vma.
1513                          */
1514                         pgoff = addr >> PAGE_SHIFT;
1515                         break;
1516                 default:
1517                         return -EINVAL;
1518                 }
1519         }
1520
1521         /*
1522          * Set 'VM_NORESERVE' if we should not account for the
1523          * memory use of this mapping.
1524          */
1525         if (flags & MAP_NORESERVE) {
1526                 /* We honor MAP_NORESERVE if allowed to overcommit */
1527                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1528                         vm_flags |= VM_NORESERVE;
1529
1530                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1531                 if (file && is_file_hugepages(file))
1532                         vm_flags |= VM_NORESERVE;
1533         }
1534
1535         addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1536         if (!IS_ERR_VALUE(addr) &&
1537             ((vm_flags & VM_LOCKED) ||
1538              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1539                 *populate = len;
1540         return addr;
1541 }
1542
1543 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1544                               unsigned long prot, unsigned long flags,
1545                               unsigned long fd, unsigned long pgoff)
1546 {
1547         struct file *file = NULL;
1548         unsigned long retval;
1549
1550         if (!(flags & MAP_ANONYMOUS)) {
1551                 audit_mmap_fd(fd, flags);
1552                 file = fget(fd);
1553                 if (!file)
1554                         return -EBADF;
1555                 if (is_file_hugepages(file))
1556                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1557                 retval = -EINVAL;
1558                 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1559                         goto out_fput;
1560         } else if (flags & MAP_HUGETLB) {
1561                 struct user_struct *user = NULL;
1562                 struct hstate *hs;
1563
1564                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1565                 if (!hs)
1566                         return -EINVAL;
1567
1568                 len = ALIGN(len, huge_page_size(hs));
1569                 /*
1570                  * VM_NORESERVE is used because the reservations will be
1571                  * taken when vm_ops->mmap() is called
1572                  * A dummy user value is used because we are not locking
1573                  * memory so no accounting is necessary
1574                  */
1575                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1576                                 VM_NORESERVE,
1577                                 &user, HUGETLB_ANONHUGE_INODE,
1578                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1579                 if (IS_ERR(file))
1580                         return PTR_ERR(file);
1581         }
1582
1583         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1584
1585         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1586 out_fput:
1587         if (file)
1588                 fput(file);
1589         return retval;
1590 }
1591
1592 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1593                 unsigned long, prot, unsigned long, flags,
1594                 unsigned long, fd, unsigned long, pgoff)
1595 {
1596         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1597 }
1598
1599 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1600 struct mmap_arg_struct {
1601         unsigned long addr;
1602         unsigned long len;
1603         unsigned long prot;
1604         unsigned long flags;
1605         unsigned long fd;
1606         unsigned long offset;
1607 };
1608
1609 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1610 {
1611         struct mmap_arg_struct a;
1612
1613         if (copy_from_user(&a, arg, sizeof(a)))
1614                 return -EFAULT;
1615         if (offset_in_page(a.offset))
1616                 return -EINVAL;
1617
1618         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1619                                a.offset >> PAGE_SHIFT);
1620 }
1621 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1622
1623 /*
1624  * Some shared mappigns will want the pages marked read-only
1625  * to track write events. If so, we'll downgrade vm_page_prot
1626  * to the private version (using protection_map[] without the
1627  * VM_SHARED bit).
1628  */
1629 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1630 {
1631         vm_flags_t vm_flags = vma->vm_flags;
1632         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1633
1634         /* If it was private or non-writable, the write bit is already clear */
1635         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1636                 return 0;
1637
1638         /* The backer wishes to know when pages are first written to? */
1639         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1640                 return 1;
1641
1642         /* The open routine did something to the protections that pgprot_modify
1643          * won't preserve? */
1644         if (pgprot_val(vm_page_prot) !=
1645             pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1646                 return 0;
1647
1648         /* Do we need to track softdirty? */
1649         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1650                 return 1;
1651
1652         /* Specialty mapping? */
1653         if (vm_flags & VM_PFNMAP)
1654                 return 0;
1655
1656         /* Can the mapping track the dirty pages? */
1657         return vma->vm_file && vma->vm_file->f_mapping &&
1658                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1659 }
1660
1661 /*
1662  * We account for memory if it's a private writeable mapping,
1663  * not hugepages and VM_NORESERVE wasn't set.
1664  */
1665 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1666 {
1667         /*
1668          * hugetlb has its own accounting separate from the core VM
1669          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1670          */
1671         if (file && is_file_hugepages(file))
1672                 return 0;
1673
1674         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1675 }
1676
1677 unsigned long mmap_region(struct file *file, unsigned long addr,
1678                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1679                 struct list_head *uf)
1680 {
1681         struct mm_struct *mm = current->mm;
1682         struct vm_area_struct *vma, *prev;
1683         int error;
1684         struct rb_node **rb_link, *rb_parent;
1685         unsigned long charged = 0;
1686
1687         /* Check against address space limit. */
1688         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1689                 unsigned long nr_pages;
1690
1691                 /*
1692                  * MAP_FIXED may remove pages of mappings that intersects with
1693                  * requested mapping. Account for the pages it would unmap.
1694                  */
1695                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1696
1697                 if (!may_expand_vm(mm, vm_flags,
1698                                         (len >> PAGE_SHIFT) - nr_pages))
1699                         return -ENOMEM;
1700         }
1701
1702         /* Clear old maps */
1703         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1704                               &rb_parent)) {
1705                 if (do_munmap(mm, addr, len, uf))
1706                         return -ENOMEM;
1707         }
1708
1709         /*
1710          * Private writable mapping: check memory availability
1711          */
1712         if (accountable_mapping(file, vm_flags)) {
1713                 charged = len >> PAGE_SHIFT;
1714                 if (security_vm_enough_memory_mm(mm, charged))
1715                         return -ENOMEM;
1716                 vm_flags |= VM_ACCOUNT;
1717         }
1718
1719         /*
1720          * Can we just expand an old mapping?
1721          */
1722         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1723                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1724         if (vma)
1725                 goto out;
1726
1727         /*
1728          * Determine the object being mapped and call the appropriate
1729          * specific mapper. the address has already been validated, but
1730          * not unmapped, but the maps are removed from the list.
1731          */
1732         vma = vm_area_alloc(mm);
1733         if (!vma) {
1734                 error = -ENOMEM;
1735                 goto unacct_error;
1736         }
1737
1738         vma->vm_start = addr;
1739         vma->vm_end = addr + len;
1740         vma->vm_flags = vm_flags;
1741         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1742         vma->vm_pgoff = pgoff;
1743
1744         if (file) {
1745                 if (vm_flags & VM_DENYWRITE) {
1746                         error = deny_write_access(file);
1747                         if (error)
1748                                 goto free_vma;
1749                 }
1750                 if (vm_flags & VM_SHARED) {
1751                         error = mapping_map_writable(file->f_mapping);
1752                         if (error)
1753                                 goto allow_write_and_free_vma;
1754                 }
1755
1756                 /* ->mmap() can change vma->vm_file, but must guarantee that
1757                  * vma_link() below can deny write-access if VM_DENYWRITE is set
1758                  * and map writably if VM_SHARED is set. This usually means the
1759                  * new file must not have been exposed to user-space, yet.
1760                  */
1761                 vma->vm_file = get_file(file);
1762                 error = call_mmap(file, vma);
1763                 if (error)
1764                         goto unmap_and_free_vma;
1765
1766                 /* Can addr have changed??
1767                  *
1768                  * Answer: Yes, several device drivers can do it in their
1769                  *         f_op->mmap method. -DaveM
1770                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1771                  *      be updated for vma_link()
1772                  */
1773                 WARN_ON_ONCE(addr != vma->vm_start);
1774
1775                 addr = vma->vm_start;
1776                 vm_flags = vma->vm_flags;
1777         } else if (vm_flags & VM_SHARED) {
1778                 error = shmem_zero_setup(vma);
1779                 if (error)
1780                         goto free_vma;
1781         } else {
1782                 vma_set_anonymous(vma);
1783         }
1784
1785         vma_link(mm, vma, prev, rb_link, rb_parent);
1786         /* Once vma denies write, undo our temporary denial count */
1787         if (file) {
1788                 if (vm_flags & VM_SHARED)
1789                         mapping_unmap_writable(file->f_mapping);
1790                 if (vm_flags & VM_DENYWRITE)
1791                         allow_write_access(file);
1792         }
1793         file = vma->vm_file;
1794 out:
1795         perf_event_mmap(vma);
1796
1797         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1798         if (vm_flags & VM_LOCKED) {
1799                 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1800                                         is_vm_hugetlb_page(vma) ||
1801                                         vma == get_gate_vma(current->mm))
1802                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1803                 else
1804                         mm->locked_vm += (len >> PAGE_SHIFT);
1805         }
1806
1807         if (file)
1808                 uprobe_mmap(vma);
1809
1810         /*
1811          * New (or expanded) vma always get soft dirty status.
1812          * Otherwise user-space soft-dirty page tracker won't
1813          * be able to distinguish situation when vma area unmapped,
1814          * then new mapped in-place (which must be aimed as
1815          * a completely new data area).
1816          */
1817         vma->vm_flags |= VM_SOFTDIRTY;
1818
1819         vma_set_page_prot(vma);
1820
1821         return addr;
1822
1823 unmap_and_free_vma:
1824         vma->vm_file = NULL;
1825         fput(file);
1826
1827         /* Undo any partial mapping done by a device driver. */
1828         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1829         charged = 0;
1830         if (vm_flags & VM_SHARED)
1831                 mapping_unmap_writable(file->f_mapping);
1832 allow_write_and_free_vma:
1833         if (vm_flags & VM_DENYWRITE)
1834                 allow_write_access(file);
1835 free_vma:
1836         vm_area_free(vma);
1837 unacct_error:
1838         if (charged)
1839                 vm_unacct_memory(charged);
1840         return error;
1841 }
1842
1843 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1844 {
1845         /*
1846          * We implement the search by looking for an rbtree node that
1847          * immediately follows a suitable gap. That is,
1848          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1849          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1850          * - gap_end - gap_start >= length
1851          */
1852
1853         struct mm_struct *mm = current->mm;
1854         struct vm_area_struct *vma;
1855         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1856
1857         /* Adjust search length to account for worst case alignment overhead */
1858         length = info->length + info->align_mask;
1859         if (length < info->length)
1860                 return -ENOMEM;
1861
1862         /* Adjust search limits by the desired length */
1863         if (info->high_limit < length)
1864                 return -ENOMEM;
1865         high_limit = info->high_limit - length;
1866
1867         if (info->low_limit > high_limit)
1868                 return -ENOMEM;
1869         low_limit = info->low_limit + length;
1870
1871         /* Check if rbtree root looks promising */
1872         if (RB_EMPTY_ROOT(&mm->mm_rb))
1873                 goto check_highest;
1874         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1875         if (vma->rb_subtree_gap < length)
1876                 goto check_highest;
1877
1878         while (true) {
1879                 /* Visit left subtree if it looks promising */
1880                 gap_end = vm_start_gap(vma);
1881                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1882                         struct vm_area_struct *left =
1883                                 rb_entry(vma->vm_rb.rb_left,
1884                                          struct vm_area_struct, vm_rb);
1885                         if (left->rb_subtree_gap >= length) {
1886                                 vma = left;
1887                                 continue;
1888                         }
1889                 }
1890
1891                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1892 check_current:
1893                 /* Check if current node has a suitable gap */
1894                 if (gap_start > high_limit)
1895                         return -ENOMEM;
1896                 if (gap_end >= low_limit &&
1897                     gap_end > gap_start && gap_end - gap_start >= length)
1898                         goto found;
1899
1900                 /* Visit right subtree if it looks promising */
1901                 if (vma->vm_rb.rb_right) {
1902                         struct vm_area_struct *right =
1903                                 rb_entry(vma->vm_rb.rb_right,
1904                                          struct vm_area_struct, vm_rb);
1905                         if (right->rb_subtree_gap >= length) {
1906                                 vma = right;
1907                                 continue;
1908                         }
1909                 }
1910
1911                 /* Go back up the rbtree to find next candidate node */
1912                 while (true) {
1913                         struct rb_node *prev = &vma->vm_rb;
1914                         if (!rb_parent(prev))
1915                                 goto check_highest;
1916                         vma = rb_entry(rb_parent(prev),
1917                                        struct vm_area_struct, vm_rb);
1918                         if (prev == vma->vm_rb.rb_left) {
1919                                 gap_start = vm_end_gap(vma->vm_prev);
1920                                 gap_end = vm_start_gap(vma);
1921                                 goto check_current;
1922                         }
1923                 }
1924         }
1925
1926 check_highest:
1927         /* Check highest gap, which does not precede any rbtree node */
1928         gap_start = mm->highest_vm_end;
1929         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1930         if (gap_start > high_limit)
1931                 return -ENOMEM;
1932
1933 found:
1934         /* We found a suitable gap. Clip it with the original low_limit. */
1935         if (gap_start < info->low_limit)
1936                 gap_start = info->low_limit;
1937
1938         /* Adjust gap address to the desired alignment */
1939         gap_start += (info->align_offset - gap_start) & info->align_mask;
1940
1941         VM_BUG_ON(gap_start + info->length > info->high_limit);
1942         VM_BUG_ON(gap_start + info->length > gap_end);
1943         return gap_start;
1944 }
1945
1946 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1947 {
1948         struct mm_struct *mm = current->mm;
1949         struct vm_area_struct *vma;
1950         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1951
1952         /* Adjust search length to account for worst case alignment overhead */
1953         length = info->length + info->align_mask;
1954         if (length < info->length)
1955                 return -ENOMEM;
1956
1957         /*
1958          * Adjust search limits by the desired length.
1959          * See implementation comment at top of unmapped_area().
1960          */
1961         gap_end = info->high_limit;
1962         if (gap_end < length)
1963                 return -ENOMEM;
1964         high_limit = gap_end - length;
1965
1966         if (info->low_limit > high_limit)
1967                 return -ENOMEM;
1968         low_limit = info->low_limit + length;
1969
1970         /* Check highest gap, which does not precede any rbtree node */
1971         gap_start = mm->highest_vm_end;
1972         if (gap_start <= high_limit)
1973                 goto found_highest;
1974
1975         /* Check if rbtree root looks promising */
1976         if (RB_EMPTY_ROOT(&mm->mm_rb))
1977                 return -ENOMEM;
1978         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1979         if (vma->rb_subtree_gap < length)
1980                 return -ENOMEM;
1981
1982         while (true) {
1983                 /* Visit right subtree if it looks promising */
1984                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1985                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1986                         struct vm_area_struct *right =
1987                                 rb_entry(vma->vm_rb.rb_right,
1988                                          struct vm_area_struct, vm_rb);
1989                         if (right->rb_subtree_gap >= length) {
1990                                 vma = right;
1991                                 continue;
1992                         }
1993                 }
1994
1995 check_current:
1996                 /* Check if current node has a suitable gap */
1997                 gap_end = vm_start_gap(vma);
1998                 if (gap_end < low_limit)
1999                         return -ENOMEM;
2000                 if (gap_start <= high_limit &&
2001                     gap_end > gap_start && gap_end - gap_start >= length)
2002                         goto found;
2003
2004                 /* Visit left subtree if it looks promising */
2005                 if (vma->vm_rb.rb_left) {
2006                         struct vm_area_struct *left =
2007                                 rb_entry(vma->vm_rb.rb_left,
2008                                          struct vm_area_struct, vm_rb);
2009                         if (left->rb_subtree_gap >= length) {
2010                                 vma = left;
2011                                 continue;
2012                         }
2013                 }
2014
2015                 /* Go back up the rbtree to find next candidate node */
2016                 while (true) {
2017                         struct rb_node *prev = &vma->vm_rb;
2018                         if (!rb_parent(prev))
2019                                 return -ENOMEM;
2020                         vma = rb_entry(rb_parent(prev),
2021                                        struct vm_area_struct, vm_rb);
2022                         if (prev == vma->vm_rb.rb_right) {
2023                                 gap_start = vma->vm_prev ?
2024                                         vm_end_gap(vma->vm_prev) : 0;
2025                                 goto check_current;
2026                         }
2027                 }
2028         }
2029
2030 found:
2031         /* We found a suitable gap. Clip it with the original high_limit. */
2032         if (gap_end > info->high_limit)
2033                 gap_end = info->high_limit;
2034
2035 found_highest:
2036         /* Compute highest gap address at the desired alignment */
2037         gap_end -= info->length;
2038         gap_end -= (gap_end - info->align_offset) & info->align_mask;
2039
2040         VM_BUG_ON(gap_end < info->low_limit);
2041         VM_BUG_ON(gap_end < gap_start);
2042         return gap_end;
2043 }
2044
2045 /* Get an address range which is currently unmapped.
2046  * For shmat() with addr=0.
2047  *
2048  * Ugly calling convention alert:
2049  * Return value with the low bits set means error value,
2050  * ie
2051  *      if (ret & ~PAGE_MASK)
2052  *              error = ret;
2053  *
2054  * This function "knows" that -ENOMEM has the bits set.
2055  */
2056 #ifndef HAVE_ARCH_UNMAPPED_AREA
2057 unsigned long
2058 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2059                 unsigned long len, unsigned long pgoff, unsigned long flags)
2060 {
2061         struct mm_struct *mm = current->mm;
2062         struct vm_area_struct *vma, *prev;
2063         struct vm_unmapped_area_info info;
2064
2065         if (len > TASK_SIZE - mmap_min_addr)
2066                 return -ENOMEM;
2067
2068         if (flags & MAP_FIXED)
2069                 return addr;
2070
2071         if (addr) {
2072                 addr = PAGE_ALIGN(addr);
2073                 vma = find_vma_prev(mm, addr, &prev);
2074                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2075                     (!vma || addr + len <= vm_start_gap(vma)) &&
2076                     (!prev || addr >= vm_end_gap(prev)))
2077                         return addr;
2078         }
2079
2080         info.flags = 0;
2081         info.length = len;
2082         info.low_limit = mm->mmap_base;
2083         info.high_limit = TASK_SIZE;
2084         info.align_mask = 0;
2085         return vm_unmapped_area(&info);
2086 }
2087 #endif
2088
2089 /*
2090  * This mmap-allocator allocates new areas top-down from below the
2091  * stack's low limit (the base):
2092  */
2093 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2094 unsigned long
2095 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2096                           const unsigned long len, const unsigned long pgoff,
2097                           const unsigned long flags)
2098 {
2099         struct vm_area_struct *vma, *prev;
2100         struct mm_struct *mm = current->mm;
2101         unsigned long addr = addr0;
2102         struct vm_unmapped_area_info info;
2103
2104         /* requested length too big for entire address space */
2105         if (len > TASK_SIZE - mmap_min_addr)
2106                 return -ENOMEM;
2107
2108         if (flags & MAP_FIXED)
2109                 return addr;
2110
2111         /* requesting a specific address */
2112         if (addr) {
2113                 addr = PAGE_ALIGN(addr);
2114                 vma = find_vma_prev(mm, addr, &prev);
2115                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2116                                 (!vma || addr + len <= vm_start_gap(vma)) &&
2117                                 (!prev || addr >= vm_end_gap(prev)))
2118                         return addr;
2119         }
2120
2121         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2122         info.length = len;
2123         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2124         info.high_limit = mm->mmap_base;
2125         info.align_mask = 0;
2126         addr = vm_unmapped_area(&info);
2127
2128         /*
2129          * A failed mmap() very likely causes application failure,
2130          * so fall back to the bottom-up function here. This scenario
2131          * can happen with large stack limits and large mmap()
2132          * allocations.
2133          */
2134         if (offset_in_page(addr)) {
2135                 VM_BUG_ON(addr != -ENOMEM);
2136                 info.flags = 0;
2137                 info.low_limit = TASK_UNMAPPED_BASE;
2138                 info.high_limit = TASK_SIZE;
2139                 addr = vm_unmapped_area(&info);
2140         }
2141
2142         return addr;
2143 }
2144 #endif
2145
2146 unsigned long
2147 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2148                 unsigned long pgoff, unsigned long flags)
2149 {
2150         unsigned long (*get_area)(struct file *, unsigned long,
2151                                   unsigned long, unsigned long, unsigned long);
2152
2153         unsigned long error = arch_mmap_check(addr, len, flags);
2154         if (error)
2155                 return error;
2156
2157         /* Careful about overflows.. */
2158         if (len > TASK_SIZE)
2159                 return -ENOMEM;
2160
2161         get_area = current->mm->get_unmapped_area;
2162         if (file) {
2163                 if (file->f_op->get_unmapped_area)
2164                         get_area = file->f_op->get_unmapped_area;
2165         } else if (flags & MAP_SHARED) {
2166                 /*
2167                  * mmap_region() will call shmem_zero_setup() to create a file,
2168                  * so use shmem's get_unmapped_area in case it can be huge.
2169                  * do_mmap_pgoff() will clear pgoff, so match alignment.
2170                  */
2171                 pgoff = 0;
2172                 get_area = shmem_get_unmapped_area;
2173         }
2174
2175         addr = get_area(file, addr, len, pgoff, flags);
2176         if (IS_ERR_VALUE(addr))
2177                 return addr;
2178
2179         if (addr > TASK_SIZE - len)
2180                 return -ENOMEM;
2181         if (offset_in_page(addr))
2182                 return -EINVAL;
2183
2184         error = security_mmap_addr(addr);
2185         return error ? error : addr;
2186 }
2187
2188 EXPORT_SYMBOL(get_unmapped_area);
2189
2190 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2191 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2192 {
2193         struct rb_node *rb_node;
2194         struct vm_area_struct *vma;
2195
2196         /* Check the cache first. */
2197         vma = vmacache_find(mm, addr);
2198         if (likely(vma))
2199                 return vma;
2200
2201         rb_node = mm->mm_rb.rb_node;
2202
2203         while (rb_node) {
2204                 struct vm_area_struct *tmp;
2205
2206                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2207
2208                 if (tmp->vm_end > addr) {
2209                         vma = tmp;
2210                         if (tmp->vm_start <= addr)
2211                                 break;
2212                         rb_node = rb_node->rb_left;
2213                 } else
2214                         rb_node = rb_node->rb_right;
2215         }
2216
2217         if (vma)
2218                 vmacache_update(addr, vma);
2219         return vma;
2220 }
2221
2222 EXPORT_SYMBOL(find_vma);
2223
2224 /*
2225  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2226  */
2227 struct vm_area_struct *
2228 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2229                         struct vm_area_struct **pprev)
2230 {
2231         struct vm_area_struct *vma;
2232
2233         vma = find_vma(mm, addr);
2234         if (vma) {
2235                 *pprev = vma->vm_prev;
2236         } else {
2237                 struct rb_node *rb_node = mm->mm_rb.rb_node;
2238                 *pprev = NULL;
2239                 while (rb_node) {
2240                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2241                         rb_node = rb_node->rb_right;
2242                 }
2243         }
2244         return vma;
2245 }
2246
2247 /*
2248  * Verify that the stack growth is acceptable and
2249  * update accounting. This is shared with both the
2250  * grow-up and grow-down cases.
2251  */
2252 static int acct_stack_growth(struct vm_area_struct *vma,
2253                              unsigned long size, unsigned long grow)
2254 {
2255         struct mm_struct *mm = vma->vm_mm;
2256         unsigned long new_start;
2257
2258         /* address space limit tests */
2259         if (!may_expand_vm(mm, vma->vm_flags, grow))
2260                 return -ENOMEM;
2261
2262         /* Stack limit test */
2263         if (size > rlimit(RLIMIT_STACK))
2264                 return -ENOMEM;
2265
2266         /* mlock limit tests */
2267         if (vma->vm_flags & VM_LOCKED) {
2268                 unsigned long locked;
2269                 unsigned long limit;
2270                 locked = mm->locked_vm + grow;
2271                 limit = rlimit(RLIMIT_MEMLOCK);
2272                 limit >>= PAGE_SHIFT;
2273                 if (locked > limit && !capable(CAP_IPC_LOCK))
2274                         return -ENOMEM;
2275         }
2276
2277         /* Check to ensure the stack will not grow into a hugetlb-only region */
2278         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2279                         vma->vm_end - size;
2280         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2281                 return -EFAULT;
2282
2283         /*
2284          * Overcommit..  This must be the final test, as it will
2285          * update security statistics.
2286          */
2287         if (security_vm_enough_memory_mm(mm, grow))
2288                 return -ENOMEM;
2289
2290         return 0;
2291 }
2292
2293 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2294 /*
2295  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2296  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2297  */
2298 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2299 {
2300         struct mm_struct *mm = vma->vm_mm;
2301         struct vm_area_struct *next;
2302         unsigned long gap_addr;
2303         int error = 0;
2304
2305         if (!(vma->vm_flags & VM_GROWSUP))
2306                 return -EFAULT;
2307
2308         /* Guard against exceeding limits of the address space. */
2309         address &= PAGE_MASK;
2310         if (address >= (TASK_SIZE & PAGE_MASK))
2311                 return -ENOMEM;
2312         address += PAGE_SIZE;
2313
2314         /* Enforce stack_guard_gap */
2315         gap_addr = address + stack_guard_gap;
2316
2317         /* Guard against overflow */
2318         if (gap_addr < address || gap_addr > TASK_SIZE)
2319                 gap_addr = TASK_SIZE;
2320
2321         next = vma->vm_next;
2322         if (next && next->vm_start < gap_addr &&
2323                         (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2324                 if (!(next->vm_flags & VM_GROWSUP))
2325                         return -ENOMEM;
2326                 /* Check that both stack segments have the same anon_vma? */
2327         }
2328
2329         /* We must make sure the anon_vma is allocated. */
2330         if (unlikely(anon_vma_prepare(vma)))
2331                 return -ENOMEM;
2332
2333         /*
2334          * vma->vm_start/vm_end cannot change under us because the caller
2335          * is required to hold the mmap_sem in read mode.  We need the
2336          * anon_vma lock to serialize against concurrent expand_stacks.
2337          */
2338         anon_vma_lock_write(vma->anon_vma);
2339
2340         /* Somebody else might have raced and expanded it already */
2341         if (address > vma->vm_end) {
2342                 unsigned long size, grow;
2343
2344                 size = address - vma->vm_start;
2345                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2346
2347                 error = -ENOMEM;
2348                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2349                         error = acct_stack_growth(vma, size, grow);
2350                         if (!error) {
2351                                 /*
2352                                  * vma_gap_update() doesn't support concurrent
2353                                  * updates, but we only hold a shared mmap_sem
2354                                  * lock here, so we need to protect against
2355                                  * concurrent vma expansions.
2356                                  * anon_vma_lock_write() doesn't help here, as
2357                                  * we don't guarantee that all growable vmas
2358                                  * in a mm share the same root anon vma.
2359                                  * So, we reuse mm->page_table_lock to guard
2360                                  * against concurrent vma expansions.
2361                                  */
2362                                 spin_lock(&mm->page_table_lock);
2363                                 if (vma->vm_flags & VM_LOCKED)
2364                                         mm->locked_vm += grow;
2365                                 vm_stat_account(mm, vma->vm_flags, grow);
2366                                 anon_vma_interval_tree_pre_update_vma(vma);
2367                                 vma->vm_end = address;
2368                                 anon_vma_interval_tree_post_update_vma(vma);
2369                                 if (vma->vm_next)
2370                                         vma_gap_update(vma->vm_next);
2371                                 else
2372                                         mm->highest_vm_end = vm_end_gap(vma);
2373                                 spin_unlock(&mm->page_table_lock);
2374
2375                                 perf_event_mmap(vma);
2376                         }
2377                 }
2378         }
2379         anon_vma_unlock_write(vma->anon_vma);
2380         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2381         validate_mm(mm);
2382         return error;
2383 }
2384 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2385
2386 /*
2387  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2388  */
2389 int expand_downwards(struct vm_area_struct *vma,
2390                                    unsigned long address)
2391 {
2392         struct mm_struct *mm = vma->vm_mm;
2393         struct vm_area_struct *prev;
2394         int error;
2395
2396         address &= PAGE_MASK;
2397         error = security_mmap_addr(address);
2398         if (error)
2399                 return error;
2400
2401         /* Enforce stack_guard_gap */
2402         prev = vma->vm_prev;
2403         /* Check that both stack segments have the same anon_vma? */
2404         if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2405                         (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2406                 if (address - prev->vm_end < stack_guard_gap)
2407                         return -ENOMEM;
2408         }
2409
2410         /* We must make sure the anon_vma is allocated. */
2411         if (unlikely(anon_vma_prepare(vma)))
2412                 return -ENOMEM;
2413
2414         /*
2415          * vma->vm_start/vm_end cannot change under us because the caller
2416          * is required to hold the mmap_sem in read mode.  We need the
2417          * anon_vma lock to serialize against concurrent expand_stacks.
2418          */
2419         anon_vma_lock_write(vma->anon_vma);
2420
2421         /* Somebody else might have raced and expanded it already */
2422         if (address < vma->vm_start) {
2423                 unsigned long size, grow;
2424
2425                 size = vma->vm_end - address;
2426                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2427
2428                 error = -ENOMEM;
2429                 if (grow <= vma->vm_pgoff) {
2430                         error = acct_stack_growth(vma, size, grow);
2431                         if (!error) {
2432                                 /*
2433                                  * vma_gap_update() doesn't support concurrent
2434                                  * updates, but we only hold a shared mmap_sem
2435                                  * lock here, so we need to protect against
2436                                  * concurrent vma expansions.
2437                                  * anon_vma_lock_write() doesn't help here, as
2438                                  * we don't guarantee that all growable vmas
2439                                  * in a mm share the same root anon vma.
2440                                  * So, we reuse mm->page_table_lock to guard
2441                                  * against concurrent vma expansions.
2442                                  */
2443                                 spin_lock(&mm->page_table_lock);
2444                                 if (vma->vm_flags & VM_LOCKED)
2445                                         mm->locked_vm += grow;
2446                                 vm_stat_account(mm, vma->vm_flags, grow);
2447                                 anon_vma_interval_tree_pre_update_vma(vma);
2448                                 vma->vm_start = address;
2449                                 vma->vm_pgoff -= grow;
2450                                 anon_vma_interval_tree_post_update_vma(vma);
2451                                 vma_gap_update(vma);
2452                                 spin_unlock(&mm->page_table_lock);
2453
2454                                 perf_event_mmap(vma);
2455                         }
2456                 }
2457         }
2458         anon_vma_unlock_write(vma->anon_vma);
2459         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2460         validate_mm(mm);
2461         return error;
2462 }
2463
2464 /* enforced gap between the expanding stack and other mappings. */
2465 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2466
2467 static int __init cmdline_parse_stack_guard_gap(char *p)
2468 {
2469         unsigned long val;
2470         char *endptr;
2471
2472         val = simple_strtoul(p, &endptr, 10);
2473         if (!*endptr)
2474                 stack_guard_gap = val << PAGE_SHIFT;
2475
2476         return 0;
2477 }
2478 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2479
2480 #ifdef CONFIG_STACK_GROWSUP
2481 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2482 {
2483         return expand_upwards(vma, address);
2484 }
2485
2486 struct vm_area_struct *
2487 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2488 {
2489         struct vm_area_struct *vma, *prev;
2490
2491         addr &= PAGE_MASK;
2492         vma = find_vma_prev(mm, addr, &prev);
2493         if (vma && (vma->vm_start <= addr))
2494                 return vma;
2495         if (!prev || expand_stack(prev, addr))
2496                 return NULL;
2497         if (prev->vm_flags & VM_LOCKED)
2498                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2499         return prev;
2500 }
2501 #else
2502 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2503 {
2504         return expand_downwards(vma, address);
2505 }
2506
2507 struct vm_area_struct *
2508 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2509 {
2510         struct vm_area_struct *vma;
2511         unsigned long start;
2512
2513         addr &= PAGE_MASK;
2514         vma = find_vma(mm, addr);
2515         if (!vma)
2516                 return NULL;
2517         if (vma->vm_start <= addr)
2518                 return vma;
2519         if (!(vma->vm_flags & VM_GROWSDOWN))
2520                 return NULL;
2521         start = vma->vm_start;
2522         if (expand_stack(vma, addr))
2523                 return NULL;
2524         if (vma->vm_flags & VM_LOCKED)
2525                 populate_vma_page_range(vma, addr, start, NULL);
2526         return vma;
2527 }
2528 #endif
2529
2530 EXPORT_SYMBOL_GPL(find_extend_vma);
2531
2532 /*
2533  * Ok - we have the memory areas we should free on the vma list,
2534  * so release them, and do the vma updates.
2535  *
2536  * Called with the mm semaphore held.
2537  */
2538 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2539 {
2540         unsigned long nr_accounted = 0;
2541
2542         /* Update high watermark before we lower total_vm */
2543         update_hiwater_vm(mm);
2544         do {
2545                 long nrpages = vma_pages(vma);
2546
2547                 if (vma->vm_flags & VM_ACCOUNT)
2548                         nr_accounted += nrpages;
2549                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2550                 vma = remove_vma(vma);
2551         } while (vma);
2552         vm_unacct_memory(nr_accounted);
2553         validate_mm(mm);
2554 }
2555
2556 /*
2557  * Get rid of page table information in the indicated region.
2558  *
2559  * Called with the mm semaphore held.
2560  */
2561 static void unmap_region(struct mm_struct *mm,
2562                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2563                 unsigned long start, unsigned long end)
2564 {
2565         struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2566         struct mmu_gather tlb;
2567
2568         lru_add_drain();
2569         tlb_gather_mmu(&tlb, mm, start, end);
2570         update_hiwater_rss(mm);
2571         unmap_vmas(&tlb, vma, start, end);
2572         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2573                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2574         tlb_finish_mmu(&tlb, start, end);
2575 }
2576
2577 /*
2578  * Create a list of vma's touched by the unmap, removing them from the mm's
2579  * vma list as we go..
2580  */
2581 static void
2582 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2583         struct vm_area_struct *prev, unsigned long end)
2584 {
2585         struct vm_area_struct **insertion_point;
2586         struct vm_area_struct *tail_vma = NULL;
2587
2588         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2589         vma->vm_prev = NULL;
2590         do {
2591                 vma_rb_erase(vma, &mm->mm_rb);
2592                 mm->map_count--;
2593                 tail_vma = vma;
2594                 vma = vma->vm_next;
2595         } while (vma && vma->vm_start < end);
2596         *insertion_point = vma;
2597         if (vma) {
2598                 vma->vm_prev = prev;
2599                 vma_gap_update(vma);
2600         } else
2601                 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2602         tail_vma->vm_next = NULL;
2603
2604         /* Kill the cache */
2605         vmacache_invalidate(mm);
2606 }
2607
2608 /*
2609  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2610  * has already been checked or doesn't make sense to fail.
2611  */
2612 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2613                 unsigned long addr, int new_below)
2614 {
2615         struct vm_area_struct *new;
2616         int err;
2617
2618         if (vma->vm_ops && vma->vm_ops->split) {
2619                 err = vma->vm_ops->split(vma, addr);
2620                 if (err)
2621                         return err;
2622         }
2623
2624         new = vm_area_dup(vma);
2625         if (!new)
2626                 return -ENOMEM;
2627
2628         if (new_below)
2629                 new->vm_end = addr;
2630         else {
2631                 new->vm_start = addr;
2632                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2633         }
2634
2635         err = vma_dup_policy(vma, new);
2636         if (err)
2637                 goto out_free_vma;
2638
2639         err = anon_vma_clone(new, vma);
2640         if (err)
2641                 goto out_free_mpol;
2642
2643         if (new->vm_file)
2644                 get_file(new->vm_file);
2645
2646         if (new->vm_ops && new->vm_ops->open)
2647                 new->vm_ops->open(new);
2648
2649         if (new_below)
2650                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2651                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2652         else
2653                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2654
2655         /* Success. */
2656         if (!err)
2657                 return 0;
2658
2659         /* Clean everything up if vma_adjust failed. */
2660         if (new->vm_ops && new->vm_ops->close)
2661                 new->vm_ops->close(new);
2662         if (new->vm_file)
2663                 fput(new->vm_file);
2664         unlink_anon_vmas(new);
2665  out_free_mpol:
2666         mpol_put(vma_policy(new));
2667  out_free_vma:
2668         vm_area_free(new);
2669         return err;
2670 }
2671
2672 /*
2673  * Split a vma into two pieces at address 'addr', a new vma is allocated
2674  * either for the first part or the tail.
2675  */
2676 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2677               unsigned long addr, int new_below)
2678 {
2679         if (mm->map_count >= sysctl_max_map_count)
2680                 return -ENOMEM;
2681
2682         return __split_vma(mm, vma, addr, new_below);
2683 }
2684
2685 /* Munmap is split into 2 main parts -- this part which finds
2686  * what needs doing, and the areas themselves, which do the
2687  * work.  This now handles partial unmappings.
2688  * Jeremy Fitzhardinge <jeremy@goop.org>
2689  */
2690 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2691                 struct list_head *uf, bool downgrade)
2692 {
2693         unsigned long end;
2694         struct vm_area_struct *vma, *prev, *last;
2695
2696         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2697                 return -EINVAL;
2698
2699         len = PAGE_ALIGN(len);
2700         if (len == 0)
2701                 return -EINVAL;
2702
2703         /* Find the first overlapping VMA */
2704         vma = find_vma(mm, start);
2705         if (!vma)
2706                 return 0;
2707         prev = vma->vm_prev;
2708         /* we have  start < vma->vm_end  */
2709
2710         /* if it doesn't overlap, we have nothing.. */
2711         end = start + len;
2712         if (vma->vm_start >= end)
2713                 return 0;
2714
2715         /*
2716          * If we need to split any vma, do it now to save pain later.
2717          *
2718          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2719          * unmapped vm_area_struct will remain in use: so lower split_vma
2720          * places tmp vma above, and higher split_vma places tmp vma below.
2721          */
2722         if (start > vma->vm_start) {
2723                 int error;
2724
2725                 /*
2726                  * Make sure that map_count on return from munmap() will
2727                  * not exceed its limit; but let map_count go just above
2728                  * its limit temporarily, to help free resources as expected.
2729                  */
2730                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2731                         return -ENOMEM;
2732
2733                 error = __split_vma(mm, vma, start, 0);
2734                 if (error)
2735                         return error;
2736                 prev = vma;
2737         }
2738
2739         /* Does it split the last one? */
2740         last = find_vma(mm, end);
2741         if (last && end > last->vm_start) {
2742                 int error = __split_vma(mm, last, end, 1);
2743                 if (error)
2744                         return error;
2745         }
2746         vma = prev ? prev->vm_next : mm->mmap;
2747
2748         if (unlikely(uf)) {
2749                 /*
2750                  * If userfaultfd_unmap_prep returns an error the vmas
2751                  * will remain splitted, but userland will get a
2752                  * highly unexpected error anyway. This is no
2753                  * different than the case where the first of the two
2754                  * __split_vma fails, but we don't undo the first
2755                  * split, despite we could. This is unlikely enough
2756                  * failure that it's not worth optimizing it for.
2757                  */
2758                 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2759                 if (error)
2760                         return error;
2761         }
2762
2763         /*
2764          * unlock any mlock()ed ranges before detaching vmas
2765          */
2766         if (mm->locked_vm) {
2767                 struct vm_area_struct *tmp = vma;
2768                 while (tmp && tmp->vm_start < end) {
2769                         if (tmp->vm_flags & VM_LOCKED) {
2770                                 mm->locked_vm -= vma_pages(tmp);
2771                                 munlock_vma_pages_all(tmp);
2772                         }
2773
2774                         tmp = tmp->vm_next;
2775                 }
2776         }
2777
2778         /* Detach vmas from rbtree */
2779         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2780
2781         /*
2782          * mpx unmap needs to be called with mmap_sem held for write.
2783          * It is safe to call it before unmap_region().
2784          */
2785         arch_unmap(mm, vma, start, end);
2786
2787         if (downgrade)
2788                 downgrade_write(&mm->mmap_sem);
2789
2790         unmap_region(mm, vma, prev, start, end);
2791
2792         /* Fix up all other VM information */
2793         remove_vma_list(mm, vma);
2794
2795         return downgrade ? 1 : 0;
2796 }
2797
2798 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2799               struct list_head *uf)
2800 {
2801         return __do_munmap(mm, start, len, uf, false);
2802 }
2803
2804 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2805 {
2806         int ret;
2807         struct mm_struct *mm = current->mm;
2808         LIST_HEAD(uf);
2809
2810         if (down_write_killable(&mm->mmap_sem))
2811                 return -EINTR;
2812
2813         ret = __do_munmap(mm, start, len, &uf, downgrade);
2814         /*
2815          * Returning 1 indicates mmap_sem is downgraded.
2816          * But 1 is not legal return value of vm_munmap() and munmap(), reset
2817          * it to 0 before return.
2818          */
2819         if (ret == 1) {
2820                 up_read(&mm->mmap_sem);
2821                 ret = 0;
2822         } else
2823                 up_write(&mm->mmap_sem);
2824
2825         userfaultfd_unmap_complete(mm, &uf);
2826         return ret;
2827 }
2828
2829 int vm_munmap(unsigned long start, size_t len)
2830 {
2831         return __vm_munmap(start, len, false);
2832 }
2833 EXPORT_SYMBOL(vm_munmap);
2834
2835 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2836 {
2837         profile_munmap(addr);
2838         return __vm_munmap(addr, len, true);
2839 }
2840
2841
2842 /*
2843  * Emulation of deprecated remap_file_pages() syscall.
2844  */
2845 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2846                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2847 {
2848
2849         struct mm_struct *mm = current->mm;
2850         struct vm_area_struct *vma;
2851         unsigned long populate = 0;
2852         unsigned long ret = -EINVAL;
2853         struct file *file;
2854
2855         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2856                      current->comm, current->pid);
2857
2858         if (prot)
2859                 return ret;
2860         start = start & PAGE_MASK;
2861         size = size & PAGE_MASK;
2862
2863         if (start + size <= start)
2864                 return ret;
2865
2866         /* Does pgoff wrap? */
2867         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2868                 return ret;
2869
2870         if (down_write_killable(&mm->mmap_sem))
2871                 return -EINTR;
2872
2873         vma = find_vma(mm, start);
2874
2875         if (!vma || !(vma->vm_flags & VM_SHARED))
2876                 goto out;
2877
2878         if (start < vma->vm_start)
2879                 goto out;
2880
2881         if (start + size > vma->vm_end) {
2882                 struct vm_area_struct *next;
2883
2884                 for (next = vma->vm_next; next; next = next->vm_next) {
2885                         /* hole between vmas ? */
2886                         if (next->vm_start != next->vm_prev->vm_end)
2887                                 goto out;
2888
2889                         if (next->vm_file != vma->vm_file)
2890                                 goto out;
2891
2892                         if (next->vm_flags != vma->vm_flags)
2893                                 goto out;
2894
2895                         if (start + size <= next->vm_end)
2896                                 break;
2897                 }
2898
2899                 if (!next)
2900                         goto out;
2901         }
2902
2903         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2904         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2905         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2906
2907         flags &= MAP_NONBLOCK;
2908         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2909         if (vma->vm_flags & VM_LOCKED) {
2910                 struct vm_area_struct *tmp;
2911                 flags |= MAP_LOCKED;
2912
2913                 /* drop PG_Mlocked flag for over-mapped range */
2914                 for (tmp = vma; tmp->vm_start >= start + size;
2915                                 tmp = tmp->vm_next) {
2916                         /*
2917                          * Split pmd and munlock page on the border
2918                          * of the range.
2919                          */
2920                         vma_adjust_trans_huge(tmp, start, start + size, 0);
2921
2922                         munlock_vma_pages_range(tmp,
2923                                         max(tmp->vm_start, start),
2924                                         min(tmp->vm_end, start + size));
2925                 }
2926         }
2927
2928         file = get_file(vma->vm_file);
2929         ret = do_mmap_pgoff(vma->vm_file, start, size,
2930                         prot, flags, pgoff, &populate, NULL);
2931         fput(file);
2932 out:
2933         up_write(&mm->mmap_sem);
2934         if (populate)
2935                 mm_populate(ret, populate);
2936         if (!IS_ERR_VALUE(ret))
2937                 ret = 0;
2938         return ret;
2939 }
2940
2941 static inline void verify_mm_writelocked(struct mm_struct *mm)
2942 {
2943 #ifdef CONFIG_DEBUG_VM
2944         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2945                 WARN_ON(1);
2946                 up_read(&mm->mmap_sem);
2947         }
2948 #endif
2949 }
2950
2951 /*
2952  *  this is really a simplified "do_mmap".  it only handles
2953  *  anonymous maps.  eventually we may be able to do some
2954  *  brk-specific accounting here.
2955  */
2956 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
2957 {
2958         struct mm_struct *mm = current->mm;
2959         struct vm_area_struct *vma, *prev;
2960         struct rb_node **rb_link, *rb_parent;
2961         pgoff_t pgoff = addr >> PAGE_SHIFT;
2962         int error;
2963
2964         /* Until we need other flags, refuse anything except VM_EXEC. */
2965         if ((flags & (~VM_EXEC)) != 0)
2966                 return -EINVAL;
2967         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2968
2969         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2970         if (offset_in_page(error))
2971                 return error;
2972
2973         error = mlock_future_check(mm, mm->def_flags, len);
2974         if (error)
2975                 return error;
2976
2977         /*
2978          * mm->mmap_sem is required to protect against another thread
2979          * changing the mappings in case we sleep.
2980          */
2981         verify_mm_writelocked(mm);
2982
2983         /*
2984          * Clear old maps.  this also does some error checking for us
2985          */
2986         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2987                               &rb_parent)) {
2988                 if (do_munmap(mm, addr, len, uf))
2989                         return -ENOMEM;
2990         }
2991
2992         /* Check against address space limits *after* clearing old maps... */
2993         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2994                 return -ENOMEM;
2995
2996         if (mm->map_count > sysctl_max_map_count)
2997                 return -ENOMEM;
2998
2999         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3000                 return -ENOMEM;
3001
3002         /* Can we just expand an old private anonymous mapping? */
3003         vma = vma_merge(mm, prev, addr, addr + len, flags,
3004                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3005         if (vma)
3006                 goto out;
3007
3008         /*
3009          * create a vma struct for an anonymous mapping
3010          */
3011         vma = vm_area_alloc(mm);
3012         if (!vma) {
3013                 vm_unacct_memory(len >> PAGE_SHIFT);
3014                 return -ENOMEM;
3015         }
3016
3017         vma_set_anonymous(vma);
3018         vma->vm_start = addr;
3019         vma->vm_end = addr + len;
3020         vma->vm_pgoff = pgoff;
3021         vma->vm_flags = flags;
3022         vma->vm_page_prot = vm_get_page_prot(flags);
3023         vma_link(mm, vma, prev, rb_link, rb_parent);
3024 out:
3025         perf_event_mmap(vma);
3026         mm->total_vm += len >> PAGE_SHIFT;
3027         mm->data_vm += len >> PAGE_SHIFT;
3028         if (flags & VM_LOCKED)
3029                 mm->locked_vm += (len >> PAGE_SHIFT);
3030         vma->vm_flags |= VM_SOFTDIRTY;
3031         return 0;
3032 }
3033
3034 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3035 {
3036         struct mm_struct *mm = current->mm;
3037         unsigned long len;
3038         int ret;
3039         bool populate;
3040         LIST_HEAD(uf);
3041
3042         len = PAGE_ALIGN(request);
3043         if (len < request)
3044                 return -ENOMEM;
3045         if (!len)
3046                 return 0;
3047
3048         if (down_write_killable(&mm->mmap_sem))
3049                 return -EINTR;
3050
3051         ret = do_brk_flags(addr, len, flags, &uf);
3052         populate = ((mm->def_flags & VM_LOCKED) != 0);
3053         up_write(&mm->mmap_sem);
3054         userfaultfd_unmap_complete(mm, &uf);
3055         if (populate && !ret)
3056                 mm_populate(addr, len);
3057         return ret;
3058 }
3059 EXPORT_SYMBOL(vm_brk_flags);
3060
3061 int vm_brk(unsigned long addr, unsigned long len)
3062 {
3063         return vm_brk_flags(addr, len, 0);
3064 }
3065 EXPORT_SYMBOL(vm_brk);
3066
3067 /* Release all mmaps. */
3068 void exit_mmap(struct mm_struct *mm)
3069 {
3070         struct mmu_gather tlb;
3071         struct vm_area_struct *vma;
3072         unsigned long nr_accounted = 0;
3073
3074         /* mm's last user has gone, and its about to be pulled down */
3075         mmu_notifier_release(mm);
3076
3077         if (unlikely(mm_is_oom_victim(mm))) {
3078                 /*
3079                  * Manually reap the mm to free as much memory as possible.
3080                  * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3081                  * this mm from further consideration.  Taking mm->mmap_sem for
3082                  * write after setting MMF_OOM_SKIP will guarantee that the oom
3083                  * reaper will not run on this mm again after mmap_sem is
3084                  * dropped.
3085                  *
3086                  * Nothing can be holding mm->mmap_sem here and the above call
3087                  * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3088                  * __oom_reap_task_mm() will not block.
3089                  *
3090                  * This needs to be done before calling munlock_vma_pages_all(),
3091                  * which clears VM_LOCKED, otherwise the oom reaper cannot
3092                  * reliably test it.
3093                  */
3094                 (void)__oom_reap_task_mm(mm);
3095
3096                 set_bit(MMF_OOM_SKIP, &mm->flags);
3097                 down_write(&mm->mmap_sem);
3098                 up_write(&mm->mmap_sem);
3099         }
3100
3101         if (mm->locked_vm) {
3102                 vma = mm->mmap;
3103                 while (vma) {
3104                         if (vma->vm_flags & VM_LOCKED)
3105                                 munlock_vma_pages_all(vma);
3106                         vma = vma->vm_next;
3107                 }
3108         }
3109
3110         arch_exit_mmap(mm);
3111
3112         vma = mm->mmap;
3113         if (!vma)       /* Can happen if dup_mmap() received an OOM */
3114                 return;
3115
3116         lru_add_drain();
3117         flush_cache_mm(mm);
3118         tlb_gather_mmu(&tlb, mm, 0, -1);
3119         /* update_hiwater_rss(mm) here? but nobody should be looking */
3120         /* Use -1 here to ensure all VMAs in the mm are unmapped */
3121         unmap_vmas(&tlb, vma, 0, -1);
3122         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3123         tlb_finish_mmu(&tlb, 0, -1);
3124
3125         /*
3126          * Walk the list again, actually closing and freeing it,
3127          * with preemption enabled, without holding any MM locks.
3128          */
3129         while (vma) {
3130                 if (vma->vm_flags & VM_ACCOUNT)
3131                         nr_accounted += vma_pages(vma);
3132                 vma = remove_vma(vma);
3133         }
3134         vm_unacct_memory(nr_accounted);
3135 }
3136
3137 /* Insert vm structure into process list sorted by address
3138  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3139  * then i_mmap_rwsem is taken here.
3140  */
3141 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3142 {
3143         struct vm_area_struct *prev;
3144         struct rb_node **rb_link, *rb_parent;
3145
3146         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3147                            &prev, &rb_link, &rb_parent))
3148                 return -ENOMEM;
3149         if ((vma->vm_flags & VM_ACCOUNT) &&
3150              security_vm_enough_memory_mm(mm, vma_pages(vma)))
3151                 return -ENOMEM;
3152
3153         /*
3154          * The vm_pgoff of a purely anonymous vma should be irrelevant
3155          * until its first write fault, when page's anon_vma and index
3156          * are set.  But now set the vm_pgoff it will almost certainly
3157          * end up with (unless mremap moves it elsewhere before that
3158          * first wfault), so /proc/pid/maps tells a consistent story.
3159          *
3160          * By setting it to reflect the virtual start address of the
3161          * vma, merges and splits can happen in a seamless way, just
3162          * using the existing file pgoff checks and manipulations.
3163          * Similarly in do_mmap_pgoff and in do_brk.
3164          */
3165         if (vma_is_anonymous(vma)) {
3166                 BUG_ON(vma->anon_vma);
3167                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3168         }
3169
3170         vma_link(mm, vma, prev, rb_link, rb_parent);
3171         return 0;
3172 }
3173
3174 /*
3175  * Copy the vma structure to a new location in the same mm,
3176  * prior to moving page table entries, to effect an mremap move.
3177  */
3178 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3179         unsigned long addr, unsigned long len, pgoff_t pgoff,
3180         bool *need_rmap_locks)
3181 {
3182         struct vm_area_struct *vma = *vmap;
3183         unsigned long vma_start = vma->vm_start;
3184         struct mm_struct *mm = vma->vm_mm;
3185         struct vm_area_struct *new_vma, *prev;
3186         struct rb_node **rb_link, *rb_parent;
3187         bool faulted_in_anon_vma = true;
3188
3189         /*
3190          * If anonymous vma has not yet been faulted, update new pgoff
3191          * to match new location, to increase its chance of merging.
3192          */
3193         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3194                 pgoff = addr >> PAGE_SHIFT;
3195                 faulted_in_anon_vma = false;
3196         }
3197
3198         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3199                 return NULL;    /* should never get here */
3200         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3201                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3202                             vma->vm_userfaultfd_ctx);
3203         if (new_vma) {
3204                 /*
3205                  * Source vma may have been merged into new_vma
3206                  */
3207                 if (unlikely(vma_start >= new_vma->vm_start &&
3208                              vma_start < new_vma->vm_end)) {
3209                         /*
3210                          * The only way we can get a vma_merge with
3211                          * self during an mremap is if the vma hasn't
3212                          * been faulted in yet and we were allowed to
3213                          * reset the dst vma->vm_pgoff to the
3214                          * destination address of the mremap to allow
3215                          * the merge to happen. mremap must change the
3216                          * vm_pgoff linearity between src and dst vmas
3217                          * (in turn preventing a vma_merge) to be
3218                          * safe. It is only safe to keep the vm_pgoff
3219                          * linear if there are no pages mapped yet.
3220                          */
3221                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3222                         *vmap = vma = new_vma;
3223                 }
3224                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3225         } else {
3226                 new_vma = vm_area_dup(vma);
3227                 if (!new_vma)
3228                         goto out;
3229                 new_vma->vm_start = addr;
3230                 new_vma->vm_end = addr + len;
3231                 new_vma->vm_pgoff = pgoff;
3232                 if (vma_dup_policy(vma, new_vma))
3233                         goto out_free_vma;
3234                 if (anon_vma_clone(new_vma, vma))
3235                         goto out_free_mempol;
3236                 if (new_vma->vm_file)
3237                         get_file(new_vma->vm_file);
3238                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3239                         new_vma->vm_ops->open(new_vma);
3240                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3241                 *need_rmap_locks = false;
3242         }
3243         return new_vma;
3244
3245 out_free_mempol:
3246         mpol_put(vma_policy(new_vma));
3247 out_free_vma:
3248         vm_area_free(new_vma);
3249 out:
3250         return NULL;
3251 }
3252
3253 /*
3254  * Return true if the calling process may expand its vm space by the passed
3255  * number of pages
3256  */
3257 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3258 {
3259         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3260                 return false;
3261
3262         if (is_data_mapping(flags) &&
3263             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3264                 /* Workaround for Valgrind */
3265                 if (rlimit(RLIMIT_DATA) == 0 &&
3266                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3267                         return true;
3268
3269                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3270                              current->comm, current->pid,
3271                              (mm->data_vm + npages) << PAGE_SHIFT,
3272                              rlimit(RLIMIT_DATA),
3273                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3274
3275                 if (!ignore_rlimit_data)
3276                         return false;
3277         }
3278
3279         return true;
3280 }
3281
3282 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3283 {
3284         mm->total_vm += npages;
3285
3286         if (is_exec_mapping(flags))
3287                 mm->exec_vm += npages;
3288         else if (is_stack_mapping(flags))
3289                 mm->stack_vm += npages;
3290         else if (is_data_mapping(flags))
3291                 mm->data_vm += npages;
3292 }
3293
3294 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3295
3296 /*
3297  * Having a close hook prevents vma merging regardless of flags.
3298  */
3299 static void special_mapping_close(struct vm_area_struct *vma)
3300 {
3301 }
3302
3303 static const char *special_mapping_name(struct vm_area_struct *vma)
3304 {
3305         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3306 }
3307
3308 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3309 {
3310         struct vm_special_mapping *sm = new_vma->vm_private_data;
3311
3312         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3313                 return -EFAULT;
3314
3315         if (sm->mremap)
3316                 return sm->mremap(sm, new_vma);
3317
3318         return 0;
3319 }
3320
3321 static const struct vm_operations_struct special_mapping_vmops = {
3322         .close = special_mapping_close,
3323         .fault = special_mapping_fault,
3324         .mremap = special_mapping_mremap,
3325         .name = special_mapping_name,
3326 };
3327
3328 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3329         .close = special_mapping_close,
3330         .fault = special_mapping_fault,
3331 };
3332
3333 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3334 {
3335         struct vm_area_struct *vma = vmf->vma;
3336         pgoff_t pgoff;
3337         struct page **pages;
3338
3339         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3340                 pages = vma->vm_private_data;
3341         } else {
3342                 struct vm_special_mapping *sm = vma->vm_private_data;
3343
3344                 if (sm->fault)
3345                         return sm->fault(sm, vmf->vma, vmf);
3346
3347                 pages = sm->pages;
3348         }
3349
3350         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3351                 pgoff--;
3352
3353         if (*pages) {
3354                 struct page *page = *pages;
3355                 get_page(page);
3356                 vmf->page = page;
3357                 return 0;
3358         }
3359
3360         return VM_FAULT_SIGBUS;
3361 }
3362
3363 static struct vm_area_struct *__install_special_mapping(
3364         struct mm_struct *mm,
3365         unsigned long addr, unsigned long len,
3366         unsigned long vm_flags, void *priv,
3367         const struct vm_operations_struct *ops)
3368 {
3369         int ret;
3370         struct vm_area_struct *vma;
3371
3372         vma = vm_area_alloc(mm);
3373         if (unlikely(vma == NULL))
3374                 return ERR_PTR(-ENOMEM);
3375
3376         vma->vm_start = addr;
3377         vma->vm_end = addr + len;
3378
3379         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3380         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3381
3382         vma->vm_ops = ops;
3383         vma->vm_private_data = priv;
3384
3385         ret = insert_vm_struct(mm, vma);
3386         if (ret)
3387                 goto out;
3388
3389         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3390
3391         perf_event_mmap(vma);
3392
3393         return vma;
3394
3395 out:
3396         vm_area_free(vma);
3397         return ERR_PTR(ret);
3398 }
3399
3400 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3401         const struct vm_special_mapping *sm)
3402 {
3403         return vma->vm_private_data == sm &&
3404                 (vma->vm_ops == &special_mapping_vmops ||
3405                  vma->vm_ops == &legacy_special_mapping_vmops);
3406 }
3407
3408 /*
3409  * Called with mm->mmap_sem held for writing.
3410  * Insert a new vma covering the given region, with the given flags.
3411  * Its pages are supplied by the given array of struct page *.
3412  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3413  * The region past the last page supplied will always produce SIGBUS.
3414  * The array pointer and the pages it points to are assumed to stay alive
3415  * for as long as this mapping might exist.
3416  */
3417 struct vm_area_struct *_install_special_mapping(
3418         struct mm_struct *mm,
3419         unsigned long addr, unsigned long len,
3420         unsigned long vm_flags, const struct vm_special_mapping *spec)
3421 {
3422         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3423                                         &special_mapping_vmops);
3424 }
3425
3426 int install_special_mapping(struct mm_struct *mm,
3427                             unsigned long addr, unsigned long len,
3428                             unsigned long vm_flags, struct page **pages)
3429 {
3430         struct vm_area_struct *vma = __install_special_mapping(
3431                 mm, addr, len, vm_flags, (void *)pages,
3432                 &legacy_special_mapping_vmops);
3433
3434         return PTR_ERR_OR_ZERO(vma);
3435 }
3436
3437 static DEFINE_MUTEX(mm_all_locks_mutex);
3438
3439 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3440 {
3441         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3442                 /*
3443                  * The LSB of head.next can't change from under us
3444                  * because we hold the mm_all_locks_mutex.
3445                  */
3446                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3447                 /*
3448                  * We can safely modify head.next after taking the
3449                  * anon_vma->root->rwsem. If some other vma in this mm shares
3450                  * the same anon_vma we won't take it again.
3451                  *
3452                  * No need of atomic instructions here, head.next
3453                  * can't change from under us thanks to the
3454                  * anon_vma->root->rwsem.
3455                  */
3456                 if (__test_and_set_bit(0, (unsigned long *)
3457                                        &anon_vma->root->rb_root.rb_root.rb_node))
3458                         BUG();
3459         }
3460 }
3461
3462 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3463 {
3464         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3465                 /*
3466                  * AS_MM_ALL_LOCKS can't change from under us because
3467                 &n