Merge tag '5.20-rc-smb3-client-fixes-part2' of git://git.samba.org/sfrench/cifs-2.6
[platform/kernel/linux-starfive.git] / mm / mmap.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * mm/mmap.c
4  *
5  * Written by obz.
6  *
7  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
8  */
9
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/backing-dev.h>
15 #include <linux/mm.h>
16 #include <linux/mm_inline.h>
17 #include <linux/vmacache.h>
18 #include <linux/shm.h>
19 #include <linux/mman.h>
20 #include <linux/pagemap.h>
21 #include <linux/swap.h>
22 #include <linux/syscalls.h>
23 #include <linux/capability.h>
24 #include <linux/init.h>
25 #include <linux/file.h>
26 #include <linux/fs.h>
27 #include <linux/personality.h>
28 #include <linux/security.h>
29 #include <linux/hugetlb.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/profile.h>
32 #include <linux/export.h>
33 #include <linux/mount.h>
34 #include <linux/mempolicy.h>
35 #include <linux/rmap.h>
36 #include <linux/mmu_notifier.h>
37 #include <linux/mmdebug.h>
38 #include <linux/perf_event.h>
39 #include <linux/audit.h>
40 #include <linux/khugepaged.h>
41 #include <linux/uprobes.h>
42 #include <linux/rbtree_augmented.h>
43 #include <linux/notifier.h>
44 #include <linux/memory.h>
45 #include <linux/printk.h>
46 #include <linux/userfaultfd_k.h>
47 #include <linux/moduleparam.h>
48 #include <linux/pkeys.h>
49 #include <linux/oom.h>
50 #include <linux/sched/mm.h>
51
52 #include <linux/uaccess.h>
53 #include <asm/cacheflush.h>
54 #include <asm/tlb.h>
55 #include <asm/mmu_context.h>
56
57 #define CREATE_TRACE_POINTS
58 #include <trace/events/mmap.h>
59
60 #include "internal.h"
61
62 #ifndef arch_mmap_check
63 #define arch_mmap_check(addr, len, flags)       (0)
64 #endif
65
66 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
67 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
68 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
69 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
70 #endif
71 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
72 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
73 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
74 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
75 #endif
76
77 static bool ignore_rlimit_data;
78 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
79
80 static void unmap_region(struct mm_struct *mm,
81                 struct vm_area_struct *vma, struct vm_area_struct *prev,
82                 unsigned long start, unsigned long end);
83
84 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
85 {
86         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
87 }
88
89 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
90 void vma_set_page_prot(struct vm_area_struct *vma)
91 {
92         unsigned long vm_flags = vma->vm_flags;
93         pgprot_t vm_page_prot;
94
95         vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
96         if (vma_wants_writenotify(vma, vm_page_prot)) {
97                 vm_flags &= ~VM_SHARED;
98                 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
99         }
100         /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
101         WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
102 }
103
104 /*
105  * Requires inode->i_mapping->i_mmap_rwsem
106  */
107 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
108                 struct file *file, struct address_space *mapping)
109 {
110         if (vma->vm_flags & VM_SHARED)
111                 mapping_unmap_writable(mapping);
112
113         flush_dcache_mmap_lock(mapping);
114         vma_interval_tree_remove(vma, &mapping->i_mmap);
115         flush_dcache_mmap_unlock(mapping);
116 }
117
118 /*
119  * Unlink a file-based vm structure from its interval tree, to hide
120  * vma from rmap and vmtruncate before freeing its page tables.
121  */
122 void unlink_file_vma(struct vm_area_struct *vma)
123 {
124         struct file *file = vma->vm_file;
125
126         if (file) {
127                 struct address_space *mapping = file->f_mapping;
128                 i_mmap_lock_write(mapping);
129                 __remove_shared_vm_struct(vma, file, mapping);
130                 i_mmap_unlock_write(mapping);
131         }
132 }
133
134 /*
135  * Close a vm structure and free it, returning the next.
136  */
137 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
138 {
139         struct vm_area_struct *next = vma->vm_next;
140
141         might_sleep();
142         if (vma->vm_ops && vma->vm_ops->close)
143                 vma->vm_ops->close(vma);
144         if (vma->vm_file)
145                 fput(vma->vm_file);
146         mpol_put(vma_policy(vma));
147         vm_area_free(vma);
148         return next;
149 }
150
151 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
152                 struct list_head *uf);
153 SYSCALL_DEFINE1(brk, unsigned long, brk)
154 {
155         unsigned long newbrk, oldbrk, origbrk;
156         struct mm_struct *mm = current->mm;
157         struct vm_area_struct *next;
158         unsigned long min_brk;
159         bool populate;
160         bool downgraded = false;
161         LIST_HEAD(uf);
162
163         if (mmap_write_lock_killable(mm))
164                 return -EINTR;
165
166         origbrk = mm->brk;
167
168 #ifdef CONFIG_COMPAT_BRK
169         /*
170          * CONFIG_COMPAT_BRK can still be overridden by setting
171          * randomize_va_space to 2, which will still cause mm->start_brk
172          * to be arbitrarily shifted
173          */
174         if (current->brk_randomized)
175                 min_brk = mm->start_brk;
176         else
177                 min_brk = mm->end_data;
178 #else
179         min_brk = mm->start_brk;
180 #endif
181         if (brk < min_brk)
182                 goto out;
183
184         /*
185          * Check against rlimit here. If this check is done later after the test
186          * of oldbrk with newbrk then it can escape the test and let the data
187          * segment grow beyond its set limit the in case where the limit is
188          * not page aligned -Ram Gupta
189          */
190         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
191                               mm->end_data, mm->start_data))
192                 goto out;
193
194         newbrk = PAGE_ALIGN(brk);
195         oldbrk = PAGE_ALIGN(mm->brk);
196         if (oldbrk == newbrk) {
197                 mm->brk = brk;
198                 goto success;
199         }
200
201         /*
202          * Always allow shrinking brk.
203          * __do_munmap() may downgrade mmap_lock to read.
204          */
205         if (brk <= mm->brk) {
206                 int ret;
207
208                 /*
209                  * mm->brk must to be protected by write mmap_lock so update it
210                  * before downgrading mmap_lock. When __do_munmap() fails,
211                  * mm->brk will be restored from origbrk.
212                  */
213                 mm->brk = brk;
214                 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
215                 if (ret < 0) {
216                         mm->brk = origbrk;
217                         goto out;
218                 } else if (ret == 1) {
219                         downgraded = true;
220                 }
221                 goto success;
222         }
223
224         /* Check against existing mmap mappings. */
225         next = find_vma(mm, oldbrk);
226         if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
227                 goto out;
228
229         /* Ok, looks good - let it rip. */
230         if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
231                 goto out;
232         mm->brk = brk;
233
234 success:
235         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
236         if (downgraded)
237                 mmap_read_unlock(mm);
238         else
239                 mmap_write_unlock(mm);
240         userfaultfd_unmap_complete(mm, &uf);
241         if (populate)
242                 mm_populate(oldbrk, newbrk - oldbrk);
243         return brk;
244
245 out:
246         mmap_write_unlock(mm);
247         return origbrk;
248 }
249
250 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
251 {
252         unsigned long gap, prev_end;
253
254         /*
255          * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
256          * allow two stack_guard_gaps between them here, and when choosing
257          * an unmapped area; whereas when expanding we only require one.
258          * That's a little inconsistent, but keeps the code here simpler.
259          */
260         gap = vm_start_gap(vma);
261         if (vma->vm_prev) {
262                 prev_end = vm_end_gap(vma->vm_prev);
263                 if (gap > prev_end)
264                         gap -= prev_end;
265                 else
266                         gap = 0;
267         }
268         return gap;
269 }
270
271 #ifdef CONFIG_DEBUG_VM_RB
272 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
273 {
274         unsigned long max = vma_compute_gap(vma), subtree_gap;
275         if (vma->vm_rb.rb_left) {
276                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
277                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
278                 if (subtree_gap > max)
279                         max = subtree_gap;
280         }
281         if (vma->vm_rb.rb_right) {
282                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
283                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
284                 if (subtree_gap > max)
285                         max = subtree_gap;
286         }
287         return max;
288 }
289
290 static int browse_rb(struct mm_struct *mm)
291 {
292         struct rb_root *root = &mm->mm_rb;
293         int i = 0, j, bug = 0;
294         struct rb_node *nd, *pn = NULL;
295         unsigned long prev = 0, pend = 0;
296
297         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
298                 struct vm_area_struct *vma;
299                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
300                 if (vma->vm_start < prev) {
301                         pr_emerg("vm_start %lx < prev %lx\n",
302                                   vma->vm_start, prev);
303                         bug = 1;
304                 }
305                 if (vma->vm_start < pend) {
306                         pr_emerg("vm_start %lx < pend %lx\n",
307                                   vma->vm_start, pend);
308                         bug = 1;
309                 }
310                 if (vma->vm_start > vma->vm_end) {
311                         pr_emerg("vm_start %lx > vm_end %lx\n",
312                                   vma->vm_start, vma->vm_end);
313                         bug = 1;
314                 }
315                 spin_lock(&mm->page_table_lock);
316                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
317                         pr_emerg("free gap %lx, correct %lx\n",
318                                vma->rb_subtree_gap,
319                                vma_compute_subtree_gap(vma));
320                         bug = 1;
321                 }
322                 spin_unlock(&mm->page_table_lock);
323                 i++;
324                 pn = nd;
325                 prev = vma->vm_start;
326                 pend = vma->vm_end;
327         }
328         j = 0;
329         for (nd = pn; nd; nd = rb_prev(nd))
330                 j++;
331         if (i != j) {
332                 pr_emerg("backwards %d, forwards %d\n", j, i);
333                 bug = 1;
334         }
335         return bug ? -1 : i;
336 }
337
338 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
339 {
340         struct rb_node *nd;
341
342         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
343                 struct vm_area_struct *vma;
344                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
345                 VM_BUG_ON_VMA(vma != ignore &&
346                         vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
347                         vma);
348         }
349 }
350
351 static void validate_mm(struct mm_struct *mm)
352 {
353         int bug = 0;
354         int i = 0;
355         unsigned long highest_address = 0;
356         struct vm_area_struct *vma = mm->mmap;
357
358         while (vma) {
359                 struct anon_vma *anon_vma = vma->anon_vma;
360                 struct anon_vma_chain *avc;
361
362                 if (anon_vma) {
363                         anon_vma_lock_read(anon_vma);
364                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
365                                 anon_vma_interval_tree_verify(avc);
366                         anon_vma_unlock_read(anon_vma);
367                 }
368
369                 highest_address = vm_end_gap(vma);
370                 vma = vma->vm_next;
371                 i++;
372         }
373         if (i != mm->map_count) {
374                 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
375                 bug = 1;
376         }
377         if (highest_address != mm->highest_vm_end) {
378                 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
379                           mm->highest_vm_end, highest_address);
380                 bug = 1;
381         }
382         i = browse_rb(mm);
383         if (i != mm->map_count) {
384                 if (i != -1)
385                         pr_emerg("map_count %d rb %d\n", mm->map_count, i);
386                 bug = 1;
387         }
388         VM_BUG_ON_MM(bug, mm);
389 }
390 #else
391 #define validate_mm_rb(root, ignore) do { } while (0)
392 #define validate_mm(mm) do { } while (0)
393 #endif
394
395 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
396                          struct vm_area_struct, vm_rb,
397                          unsigned long, rb_subtree_gap, vma_compute_gap)
398
399 /*
400  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
401  * vma->vm_prev->vm_end values changed, without modifying the vma's position
402  * in the rbtree.
403  */
404 static void vma_gap_update(struct vm_area_struct *vma)
405 {
406         /*
407          * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
408          * a callback function that does exactly what we want.
409          */
410         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
411 }
412
413 static inline void vma_rb_insert(struct vm_area_struct *vma,
414                                  struct rb_root *root)
415 {
416         /* All rb_subtree_gap values must be consistent prior to insertion */
417         validate_mm_rb(root, NULL);
418
419         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
420 }
421
422 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
423 {
424         /*
425          * Note rb_erase_augmented is a fairly large inline function,
426          * so make sure we instantiate it only once with our desired
427          * augmented rbtree callbacks.
428          */
429         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
430 }
431
432 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
433                                                 struct rb_root *root,
434                                                 struct vm_area_struct *ignore)
435 {
436         /*
437          * All rb_subtree_gap values must be consistent prior to erase,
438          * with the possible exception of
439          *
440          * a. the "next" vma being erased if next->vm_start was reduced in
441          *    __vma_adjust() -> __vma_unlink()
442          * b. the vma being erased in detach_vmas_to_be_unmapped() ->
443          *    vma_rb_erase()
444          */
445         validate_mm_rb(root, ignore);
446
447         __vma_rb_erase(vma, root);
448 }
449
450 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
451                                          struct rb_root *root)
452 {
453         vma_rb_erase_ignore(vma, root, vma);
454 }
455
456 /*
457  * vma has some anon_vma assigned, and is already inserted on that
458  * anon_vma's interval trees.
459  *
460  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
461  * vma must be removed from the anon_vma's interval trees using
462  * anon_vma_interval_tree_pre_update_vma().
463  *
464  * After the update, the vma will be reinserted using
465  * anon_vma_interval_tree_post_update_vma().
466  *
467  * The entire update must be protected by exclusive mmap_lock and by
468  * the root anon_vma's mutex.
469  */
470 static inline void
471 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
472 {
473         struct anon_vma_chain *avc;
474
475         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
476                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
477 }
478
479 static inline void
480 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
481 {
482         struct anon_vma_chain *avc;
483
484         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
485                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
486 }
487
488 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
489                 unsigned long end, struct vm_area_struct **pprev,
490                 struct rb_node ***rb_link, struct rb_node **rb_parent)
491 {
492         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
493
494         mmap_assert_locked(mm);
495         __rb_link = &mm->mm_rb.rb_node;
496         rb_prev = __rb_parent = NULL;
497
498         while (*__rb_link) {
499                 struct vm_area_struct *vma_tmp;
500
501                 __rb_parent = *__rb_link;
502                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
503
504                 if (vma_tmp->vm_end > addr) {
505                         /* Fail if an existing vma overlaps the area */
506                         if (vma_tmp->vm_start < end)
507                                 return -ENOMEM;
508                         __rb_link = &__rb_parent->rb_left;
509                 } else {
510                         rb_prev = __rb_parent;
511                         __rb_link = &__rb_parent->rb_right;
512                 }
513         }
514
515         *pprev = NULL;
516         if (rb_prev)
517                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
518         *rb_link = __rb_link;
519         *rb_parent = __rb_parent;
520         return 0;
521 }
522
523 /*
524  * vma_next() - Get the next VMA.
525  * @mm: The mm_struct.
526  * @vma: The current vma.
527  *
528  * If @vma is NULL, return the first vma in the mm.
529  *
530  * Returns: The next VMA after @vma.
531  */
532 static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
533                                          struct vm_area_struct *vma)
534 {
535         if (!vma)
536                 return mm->mmap;
537
538         return vma->vm_next;
539 }
540
541 /*
542  * munmap_vma_range() - munmap VMAs that overlap a range.
543  * @mm: The mm struct
544  * @start: The start of the range.
545  * @len: The length of the range.
546  * @pprev: pointer to the pointer that will be set to previous vm_area_struct
547  * @rb_link: the rb_node
548  * @rb_parent: the parent rb_node
549  *
550  * Find all the vm_area_struct that overlap from @start to
551  * @end and munmap them.  Set @pprev to the previous vm_area_struct.
552  *
553  * Returns: -ENOMEM on munmap failure or 0 on success.
554  */
555 static inline int
556 munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
557                  struct vm_area_struct **pprev, struct rb_node ***link,
558                  struct rb_node **parent, struct list_head *uf)
559 {
560
561         while (find_vma_links(mm, start, start + len, pprev, link, parent))
562                 if (do_munmap(mm, start, len, uf))
563                         return -ENOMEM;
564
565         return 0;
566 }
567 static unsigned long count_vma_pages_range(struct mm_struct *mm,
568                 unsigned long addr, unsigned long end)
569 {
570         unsigned long nr_pages = 0;
571         struct vm_area_struct *vma;
572
573         /* Find first overlapping mapping */
574         vma = find_vma_intersection(mm, addr, end);
575         if (!vma)
576                 return 0;
577
578         nr_pages = (min(end, vma->vm_end) -
579                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
580
581         /* Iterate over the rest of the overlaps */
582         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
583                 unsigned long overlap_len;
584
585                 if (vma->vm_start > end)
586                         break;
587
588                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
589                 nr_pages += overlap_len >> PAGE_SHIFT;
590         }
591
592         return nr_pages;
593 }
594
595 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
596                 struct rb_node **rb_link, struct rb_node *rb_parent)
597 {
598         /* Update tracking information for the gap following the new vma. */
599         if (vma->vm_next)
600                 vma_gap_update(vma->vm_next);
601         else
602                 mm->highest_vm_end = vm_end_gap(vma);
603
604         /*
605          * vma->vm_prev wasn't known when we followed the rbtree to find the
606          * correct insertion point for that vma. As a result, we could not
607          * update the vma vm_rb parents rb_subtree_gap values on the way down.
608          * So, we first insert the vma with a zero rb_subtree_gap value
609          * (to be consistent with what we did on the way down), and then
610          * immediately update the gap to the correct value. Finally we
611          * rebalance the rbtree after all augmented values have been set.
612          */
613         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
614         vma->rb_subtree_gap = 0;
615         vma_gap_update(vma);
616         vma_rb_insert(vma, &mm->mm_rb);
617 }
618
619 static void __vma_link_file(struct vm_area_struct *vma)
620 {
621         struct file *file;
622
623         file = vma->vm_file;
624         if (file) {
625                 struct address_space *mapping = file->f_mapping;
626
627                 if (vma->vm_flags & VM_SHARED)
628                         mapping_allow_writable(mapping);
629
630                 flush_dcache_mmap_lock(mapping);
631                 vma_interval_tree_insert(vma, &mapping->i_mmap);
632                 flush_dcache_mmap_unlock(mapping);
633         }
634 }
635
636 static void
637 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
638         struct vm_area_struct *prev, struct rb_node **rb_link,
639         struct rb_node *rb_parent)
640 {
641         __vma_link_list(mm, vma, prev);
642         __vma_link_rb(mm, vma, rb_link, rb_parent);
643 }
644
645 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
646                         struct vm_area_struct *prev, struct rb_node **rb_link,
647                         struct rb_node *rb_parent)
648 {
649         struct address_space *mapping = NULL;
650
651         if (vma->vm_file) {
652                 mapping = vma->vm_file->f_mapping;
653                 i_mmap_lock_write(mapping);
654         }
655
656         __vma_link(mm, vma, prev, rb_link, rb_parent);
657         __vma_link_file(vma);
658
659         if (mapping)
660                 i_mmap_unlock_write(mapping);
661
662         mm->map_count++;
663         validate_mm(mm);
664 }
665
666 /*
667  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
668  * mm's list and rbtree.  It has already been inserted into the interval tree.
669  */
670 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
671 {
672         struct vm_area_struct *prev;
673         struct rb_node **rb_link, *rb_parent;
674
675         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
676                            &prev, &rb_link, &rb_parent))
677                 BUG();
678         __vma_link(mm, vma, prev, rb_link, rb_parent);
679         mm->map_count++;
680 }
681
682 static __always_inline void __vma_unlink(struct mm_struct *mm,
683                                                 struct vm_area_struct *vma,
684                                                 struct vm_area_struct *ignore)
685 {
686         vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
687         __vma_unlink_list(mm, vma);
688         /* Kill the cache */
689         vmacache_invalidate(mm);
690 }
691
692 /*
693  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
694  * is already present in an i_mmap tree without adjusting the tree.
695  * The following helper function should be used when such adjustments
696  * are necessary.  The "insert" vma (if any) is to be inserted
697  * before we drop the necessary locks.
698  */
699 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
700         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
701         struct vm_area_struct *expand)
702 {
703         struct mm_struct *mm = vma->vm_mm;
704         struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
705         struct address_space *mapping = NULL;
706         struct rb_root_cached *root = NULL;
707         struct anon_vma *anon_vma = NULL;
708         struct file *file = vma->vm_file;
709         bool start_changed = false, end_changed = false;
710         long adjust_next = 0;
711         int remove_next = 0;
712
713         if (next && !insert) {
714                 struct vm_area_struct *exporter = NULL, *importer = NULL;
715
716                 if (end >= next->vm_end) {
717                         /*
718                          * vma expands, overlapping all the next, and
719                          * perhaps the one after too (mprotect case 6).
720                          * The only other cases that gets here are
721                          * case 1, case 7 and case 8.
722                          */
723                         if (next == expand) {
724                                 /*
725                                  * The only case where we don't expand "vma"
726                                  * and we expand "next" instead is case 8.
727                                  */
728                                 VM_WARN_ON(end != next->vm_end);
729                                 /*
730                                  * remove_next == 3 means we're
731                                  * removing "vma" and that to do so we
732                                  * swapped "vma" and "next".
733                                  */
734                                 remove_next = 3;
735                                 VM_WARN_ON(file != next->vm_file);
736                                 swap(vma, next);
737                         } else {
738                                 VM_WARN_ON(expand != vma);
739                                 /*
740                                  * case 1, 6, 7, remove_next == 2 is case 6,
741                                  * remove_next == 1 is case 1 or 7.
742                                  */
743                                 remove_next = 1 + (end > next->vm_end);
744                                 VM_WARN_ON(remove_next == 2 &&
745                                            end != next->vm_next->vm_end);
746                                 /* trim end to next, for case 6 first pass */
747                                 end = next->vm_end;
748                         }
749
750                         exporter = next;
751                         importer = vma;
752
753                         /*
754                          * If next doesn't have anon_vma, import from vma after
755                          * next, if the vma overlaps with it.
756                          */
757                         if (remove_next == 2 && !next->anon_vma)
758                                 exporter = next->vm_next;
759
760                 } else if (end > next->vm_start) {
761                         /*
762                          * vma expands, overlapping part of the next:
763                          * mprotect case 5 shifting the boundary up.
764                          */
765                         adjust_next = (end - next->vm_start);
766                         exporter = next;
767                         importer = vma;
768                         VM_WARN_ON(expand != importer);
769                 } else if (end < vma->vm_end) {
770                         /*
771                          * vma shrinks, and !insert tells it's not
772                          * split_vma inserting another: so it must be
773                          * mprotect case 4 shifting the boundary down.
774                          */
775                         adjust_next = -(vma->vm_end - end);
776                         exporter = vma;
777                         importer = next;
778                         VM_WARN_ON(expand != importer);
779                 }
780
781                 /*
782                  * Easily overlooked: when mprotect shifts the boundary,
783                  * make sure the expanding vma has anon_vma set if the
784                  * shrinking vma had, to cover any anon pages imported.
785                  */
786                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
787                         int error;
788
789                         importer->anon_vma = exporter->anon_vma;
790                         error = anon_vma_clone(importer, exporter);
791                         if (error)
792                                 return error;
793                 }
794         }
795 again:
796         vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
797
798         if (file) {
799                 mapping = file->f_mapping;
800                 root = &mapping->i_mmap;
801                 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
802
803                 if (adjust_next)
804                         uprobe_munmap(next, next->vm_start, next->vm_end);
805
806                 i_mmap_lock_write(mapping);
807                 if (insert) {
808                         /*
809                          * Put into interval tree now, so instantiated pages
810                          * are visible to arm/parisc __flush_dcache_page
811                          * throughout; but we cannot insert into address
812                          * space until vma start or end is updated.
813                          */
814                         __vma_link_file(insert);
815                 }
816         }
817
818         anon_vma = vma->anon_vma;
819         if (!anon_vma && adjust_next)
820                 anon_vma = next->anon_vma;
821         if (anon_vma) {
822                 VM_WARN_ON(adjust_next && next->anon_vma &&
823                            anon_vma != next->anon_vma);
824                 anon_vma_lock_write(anon_vma);
825                 anon_vma_interval_tree_pre_update_vma(vma);
826                 if (adjust_next)
827                         anon_vma_interval_tree_pre_update_vma(next);
828         }
829
830         if (file) {
831                 flush_dcache_mmap_lock(mapping);
832                 vma_interval_tree_remove(vma, root);
833                 if (adjust_next)
834                         vma_interval_tree_remove(next, root);
835         }
836
837         if (start != vma->vm_start) {
838                 vma->vm_start = start;
839                 start_changed = true;
840         }
841         if (end != vma->vm_end) {
842                 vma->vm_end = end;
843                 end_changed = true;
844         }
845         vma->vm_pgoff = pgoff;
846         if (adjust_next) {
847                 next->vm_start += adjust_next;
848                 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
849         }
850
851         if (file) {
852                 if (adjust_next)
853                         vma_interval_tree_insert(next, root);
854                 vma_interval_tree_insert(vma, root);
855                 flush_dcache_mmap_unlock(mapping);
856         }
857
858         if (remove_next) {
859                 /*
860                  * vma_merge has merged next into vma, and needs
861                  * us to remove next before dropping the locks.
862                  */
863                 if (remove_next != 3)
864                         __vma_unlink(mm, next, next);
865                 else
866                         /*
867                          * vma is not before next if they've been
868                          * swapped.
869                          *
870                          * pre-swap() next->vm_start was reduced so
871                          * tell validate_mm_rb to ignore pre-swap()
872                          * "next" (which is stored in post-swap()
873                          * "vma").
874                          */
875                         __vma_unlink(mm, next, vma);
876                 if (file)
877                         __remove_shared_vm_struct(next, file, mapping);
878         } else if (insert) {
879                 /*
880                  * split_vma has split insert from vma, and needs
881                  * us to insert it before dropping the locks
882                  * (it may either follow vma or precede it).
883                  */
884                 __insert_vm_struct(mm, insert);
885         } else {
886                 if (start_changed)
887                         vma_gap_update(vma);
888                 if (end_changed) {
889                         if (!next)
890                                 mm->highest_vm_end = vm_end_gap(vma);
891                         else if (!adjust_next)
892                                 vma_gap_update(next);
893                 }
894         }
895
896         if (anon_vma) {
897                 anon_vma_interval_tree_post_update_vma(vma);
898                 if (adjust_next)
899                         anon_vma_interval_tree_post_update_vma(next);
900                 anon_vma_unlock_write(anon_vma);
901         }
902
903         if (file) {
904                 i_mmap_unlock_write(mapping);
905                 uprobe_mmap(vma);
906
907                 if (adjust_next)
908                         uprobe_mmap(next);
909         }
910
911         if (remove_next) {
912                 if (file) {
913                         uprobe_munmap(next, next->vm_start, next->vm_end);
914                         fput(file);
915                 }
916                 if (next->anon_vma)
917                         anon_vma_merge(vma, next);
918                 mm->map_count--;
919                 mpol_put(vma_policy(next));
920                 vm_area_free(next);
921                 /*
922                  * In mprotect's case 6 (see comments on vma_merge),
923                  * we must remove another next too. It would clutter
924                  * up the code too much to do both in one go.
925                  */
926                 if (remove_next != 3) {
927                         /*
928                          * If "next" was removed and vma->vm_end was
929                          * expanded (up) over it, in turn
930                          * "next->vm_prev->vm_end" changed and the
931                          * "vma->vm_next" gap must be updated.
932                          */
933                         next = vma->vm_next;
934                 } else {
935                         /*
936                          * For the scope of the comment "next" and
937                          * "vma" considered pre-swap(): if "vma" was
938                          * removed, next->vm_start was expanded (down)
939                          * over it and the "next" gap must be updated.
940                          * Because of the swap() the post-swap() "vma"
941                          * actually points to pre-swap() "next"
942                          * (post-swap() "next" as opposed is now a
943                          * dangling pointer).
944                          */
945                         next = vma;
946                 }
947                 if (remove_next == 2) {
948                         remove_next = 1;
949                         end = next->vm_end;
950                         goto again;
951                 }
952                 else if (next)
953                         vma_gap_update(next);
954                 else {
955                         /*
956                          * If remove_next == 2 we obviously can't
957                          * reach this path.
958                          *
959                          * If remove_next == 3 we can't reach this
960                          * path because pre-swap() next is always not
961                          * NULL. pre-swap() "next" is not being
962                          * removed and its next->vm_end is not altered
963                          * (and furthermore "end" already matches
964                          * next->vm_end in remove_next == 3).
965                          *
966                          * We reach this only in the remove_next == 1
967                          * case if the "next" vma that was removed was
968                          * the highest vma of the mm. However in such
969                          * case next->vm_end == "end" and the extended
970                          * "vma" has vma->vm_end == next->vm_end so
971                          * mm->highest_vm_end doesn't need any update
972                          * in remove_next == 1 case.
973                          */
974                         VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
975                 }
976         }
977         if (insert && file)
978                 uprobe_mmap(insert);
979
980         validate_mm(mm);
981
982         return 0;
983 }
984
985 /*
986  * If the vma has a ->close operation then the driver probably needs to release
987  * per-vma resources, so we don't attempt to merge those.
988  */
989 static inline int is_mergeable_vma(struct vm_area_struct *vma,
990                                 struct file *file, unsigned long vm_flags,
991                                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
992                                 struct anon_vma_name *anon_name)
993 {
994         /*
995          * VM_SOFTDIRTY should not prevent from VMA merging, if we
996          * match the flags but dirty bit -- the caller should mark
997          * merged VMA as dirty. If dirty bit won't be excluded from
998          * comparison, we increase pressure on the memory system forcing
999          * the kernel to generate new VMAs when old one could be
1000          * extended instead.
1001          */
1002         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1003                 return 0;
1004         if (vma->vm_file != file)
1005                 return 0;
1006         if (vma->vm_ops && vma->vm_ops->close)
1007                 return 0;
1008         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1009                 return 0;
1010         if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
1011                 return 0;
1012         return 1;
1013 }
1014
1015 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1016                                         struct anon_vma *anon_vma2,
1017                                         struct vm_area_struct *vma)
1018 {
1019         /*
1020          * The list_is_singular() test is to avoid merging VMA cloned from
1021          * parents. This can improve scalability caused by anon_vma lock.
1022          */
1023         if ((!anon_vma1 || !anon_vma2) && (!vma ||
1024                 list_is_singular(&vma->anon_vma_chain)))
1025                 return 1;
1026         return anon_vma1 == anon_vma2;
1027 }
1028
1029 /*
1030  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1031  * in front of (at a lower virtual address and file offset than) the vma.
1032  *
1033  * We cannot merge two vmas if they have differently assigned (non-NULL)
1034  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1035  *
1036  * We don't check here for the merged mmap wrapping around the end of pagecache
1037  * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1038  * wrap, nor mmaps which cover the final page at index -1UL.
1039  */
1040 static int
1041 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1042                      struct anon_vma *anon_vma, struct file *file,
1043                      pgoff_t vm_pgoff,
1044                      struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1045                      struct anon_vma_name *anon_name)
1046 {
1047         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
1048             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1049                 if (vma->vm_pgoff == vm_pgoff)
1050                         return 1;
1051         }
1052         return 0;
1053 }
1054
1055 /*
1056  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1057  * beyond (at a higher virtual address and file offset than) the vma.
1058  *
1059  * We cannot merge two vmas if they have differently assigned (non-NULL)
1060  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1061  */
1062 static int
1063 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1064                     struct anon_vma *anon_vma, struct file *file,
1065                     pgoff_t vm_pgoff,
1066                     struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1067                     struct anon_vma_name *anon_name)
1068 {
1069         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
1070             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1071                 pgoff_t vm_pglen;
1072                 vm_pglen = vma_pages(vma);
1073                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1074                         return 1;
1075         }
1076         return 0;
1077 }
1078
1079 /*
1080  * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
1081  * figure out whether that can be merged with its predecessor or its
1082  * successor.  Or both (it neatly fills a hole).
1083  *
1084  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1085  * certain not to be mapped by the time vma_merge is called; but when
1086  * called for mprotect, it is certain to be already mapped (either at
1087  * an offset within prev, or at the start of next), and the flags of
1088  * this area are about to be changed to vm_flags - and the no-change
1089  * case has already been eliminated.
1090  *
1091  * The following mprotect cases have to be considered, where AAAA is
1092  * the area passed down from mprotect_fixup, never extending beyond one
1093  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1094  *
1095  *     AAAA             AAAA                   AAAA
1096  *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPNNNNNN
1097  *    cannot merge    might become       might become
1098  *                    PPNNNNNNNNNN       PPPPPPPPPPNN
1099  *    mmap, brk or    case 4 below       case 5 below
1100  *    mremap move:
1101  *                        AAAA               AAAA
1102  *                    PPPP    NNNN       PPPPNNNNXXXX
1103  *                    might become       might become
1104  *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
1105  *                    PPPPPPPPNNNN 2 or  PPPPPPPPXXXX 7 or
1106  *                    PPPPNNNNNNNN 3     PPPPXXXXXXXX 8
1107  *
1108  * It is important for case 8 that the vma NNNN overlapping the
1109  * region AAAA is never going to extended over XXXX. Instead XXXX must
1110  * be extended in region AAAA and NNNN must be removed. This way in
1111  * all cases where vma_merge succeeds, the moment vma_adjust drops the
1112  * rmap_locks, the properties of the merged vma will be already
1113  * correct for the whole merged range. Some of those properties like
1114  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1115  * be correct for the whole merged range immediately after the
1116  * rmap_locks are released. Otherwise if XXXX would be removed and
1117  * NNNN would be extended over the XXXX range, remove_migration_ptes
1118  * or other rmap walkers (if working on addresses beyond the "end"
1119  * parameter) may establish ptes with the wrong permissions of NNNN
1120  * instead of the right permissions of XXXX.
1121  */
1122 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1123                         struct vm_area_struct *prev, unsigned long addr,
1124                         unsigned long end, unsigned long vm_flags,
1125                         struct anon_vma *anon_vma, struct file *file,
1126                         pgoff_t pgoff, struct mempolicy *policy,
1127                         struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1128                         struct anon_vma_name *anon_name)
1129 {
1130         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1131         struct vm_area_struct *area, *next;
1132         int err;
1133
1134         /*
1135          * We later require that vma->vm_flags == vm_flags,
1136          * so this tests vma->vm_flags & VM_SPECIAL, too.
1137          */
1138         if (vm_flags & VM_SPECIAL)
1139                 return NULL;
1140
1141         next = vma_next(mm, prev);
1142         area = next;
1143         if (area && area->vm_end == end)                /* cases 6, 7, 8 */
1144                 next = next->vm_next;
1145
1146         /* verify some invariant that must be enforced by the caller */
1147         VM_WARN_ON(prev && addr <= prev->vm_start);
1148         VM_WARN_ON(area && end > area->vm_end);
1149         VM_WARN_ON(addr >= end);
1150
1151         /*
1152          * Can it merge with the predecessor?
1153          */
1154         if (prev && prev->vm_end == addr &&
1155                         mpol_equal(vma_policy(prev), policy) &&
1156                         can_vma_merge_after(prev, vm_flags,
1157                                             anon_vma, file, pgoff,
1158                                             vm_userfaultfd_ctx, anon_name)) {
1159                 /*
1160                  * OK, it can.  Can we now merge in the successor as well?
1161                  */
1162                 if (next && end == next->vm_start &&
1163                                 mpol_equal(policy, vma_policy(next)) &&
1164                                 can_vma_merge_before(next, vm_flags,
1165                                                      anon_vma, file,
1166                                                      pgoff+pglen,
1167                                                      vm_userfaultfd_ctx, anon_name) &&
1168                                 is_mergeable_anon_vma(prev->anon_vma,
1169                                                       next->anon_vma, NULL)) {
1170                                                         /* cases 1, 6 */
1171                         err = __vma_adjust(prev, prev->vm_start,
1172                                          next->vm_end, prev->vm_pgoff, NULL,
1173                                          prev);
1174                 } else                                  /* cases 2, 5, 7 */
1175                         err = __vma_adjust(prev, prev->vm_start,
1176                                          end, prev->vm_pgoff, NULL, prev);
1177                 if (err)
1178                         return NULL;
1179                 khugepaged_enter_vma(prev, vm_flags);
1180                 return prev;
1181         }
1182
1183         /*
1184          * Can this new request be merged in front of next?
1185          */
1186         if (next && end == next->vm_start &&
1187                         mpol_equal(policy, vma_policy(next)) &&
1188                         can_vma_merge_before(next, vm_flags,
1189                                              anon_vma, file, pgoff+pglen,
1190                                              vm_userfaultfd_ctx, anon_name)) {
1191                 if (prev && addr < prev->vm_end)        /* case 4 */
1192                         err = __vma_adjust(prev, prev->vm_start,
1193                                          addr, prev->vm_pgoff, NULL, next);
1194                 else {                                  /* cases 3, 8 */
1195                         err = __vma_adjust(area, addr, next->vm_end,
1196                                          next->vm_pgoff - pglen, NULL, next);
1197                         /*
1198                          * In case 3 area is already equal to next and
1199                          * this is a noop, but in case 8 "area" has
1200                          * been removed and next was expanded over it.
1201                          */
1202                         area = next;
1203                 }
1204                 if (err)
1205                         return NULL;
1206                 khugepaged_enter_vma(area, vm_flags);
1207                 return area;
1208         }
1209
1210         return NULL;
1211 }
1212
1213 /*
1214  * Rough compatibility check to quickly see if it's even worth looking
1215  * at sharing an anon_vma.
1216  *
1217  * They need to have the same vm_file, and the flags can only differ
1218  * in things that mprotect may change.
1219  *
1220  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1221  * we can merge the two vma's. For example, we refuse to merge a vma if
1222  * there is a vm_ops->close() function, because that indicates that the
1223  * driver is doing some kind of reference counting. But that doesn't
1224  * really matter for the anon_vma sharing case.
1225  */
1226 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1227 {
1228         return a->vm_end == b->vm_start &&
1229                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1230                 a->vm_file == b->vm_file &&
1231                 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1232                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1233 }
1234
1235 /*
1236  * Do some basic sanity checking to see if we can re-use the anon_vma
1237  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1238  * the same as 'old', the other will be the new one that is trying
1239  * to share the anon_vma.
1240  *
1241  * NOTE! This runs with mmap_lock held for reading, so it is possible that
1242  * the anon_vma of 'old' is concurrently in the process of being set up
1243  * by another page fault trying to merge _that_. But that's ok: if it
1244  * is being set up, that automatically means that it will be a singleton
1245  * acceptable for merging, so we can do all of this optimistically. But
1246  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1247  *
1248  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1249  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1250  * is to return an anon_vma that is "complex" due to having gone through
1251  * a fork).
1252  *
1253  * We also make sure that the two vma's are compatible (adjacent,
1254  * and with the same memory policies). That's all stable, even with just
1255  * a read lock on the mmap_lock.
1256  */
1257 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1258 {
1259         if (anon_vma_compatible(a, b)) {
1260                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1261
1262                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1263                         return anon_vma;
1264         }
1265         return NULL;
1266 }
1267
1268 /*
1269  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1270  * neighbouring vmas for a suitable anon_vma, before it goes off
1271  * to allocate a new anon_vma.  It checks because a repetitive
1272  * sequence of mprotects and faults may otherwise lead to distinct
1273  * anon_vmas being allocated, preventing vma merge in subsequent
1274  * mprotect.
1275  */
1276 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1277 {
1278         struct anon_vma *anon_vma = NULL;
1279
1280         /* Try next first. */
1281         if (vma->vm_next) {
1282                 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1283                 if (anon_vma)
1284                         return anon_vma;
1285         }
1286
1287         /* Try prev next. */
1288         if (vma->vm_prev)
1289                 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1290
1291         /*
1292          * We might reach here with anon_vma == NULL if we can't find
1293          * any reusable anon_vma.
1294          * There's no absolute need to look only at touching neighbours:
1295          * we could search further afield for "compatible" anon_vmas.
1296          * But it would probably just be a waste of time searching,
1297          * or lead to too many vmas hanging off the same anon_vma.
1298          * We're trying to allow mprotect remerging later on,
1299          * not trying to minimize memory used for anon_vmas.
1300          */
1301         return anon_vma;
1302 }
1303
1304 /*
1305  * If a hint addr is less than mmap_min_addr change hint to be as
1306  * low as possible but still greater than mmap_min_addr
1307  */
1308 static inline unsigned long round_hint_to_min(unsigned long hint)
1309 {
1310         hint &= PAGE_MASK;
1311         if (((void *)hint != NULL) &&
1312             (hint < mmap_min_addr))
1313                 return PAGE_ALIGN(mmap_min_addr);
1314         return hint;
1315 }
1316
1317 int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1318                        unsigned long len)
1319 {
1320         unsigned long locked, lock_limit;
1321
1322         /*  mlock MCL_FUTURE? */
1323         if (flags & VM_LOCKED) {
1324                 locked = len >> PAGE_SHIFT;
1325                 locked += mm->locked_vm;
1326                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1327                 lock_limit >>= PAGE_SHIFT;
1328                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1329                         return -EAGAIN;
1330         }
1331         return 0;
1332 }
1333
1334 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1335 {
1336         if (S_ISREG(inode->i_mode))
1337                 return MAX_LFS_FILESIZE;
1338
1339         if (S_ISBLK(inode->i_mode))
1340                 return MAX_LFS_FILESIZE;
1341
1342         if (S_ISSOCK(inode->i_mode))
1343                 return MAX_LFS_FILESIZE;
1344
1345         /* Special "we do even unsigned file positions" case */
1346         if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1347                 return 0;
1348
1349         /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1350         return ULONG_MAX;
1351 }
1352
1353 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1354                                 unsigned long pgoff, unsigned long len)
1355 {
1356         u64 maxsize = file_mmap_size_max(file, inode);
1357
1358         if (maxsize && len > maxsize)
1359                 return false;
1360         maxsize -= len;
1361         if (pgoff > maxsize >> PAGE_SHIFT)
1362                 return false;
1363         return true;
1364 }
1365
1366 /*
1367  * The caller must write-lock current->mm->mmap_lock.
1368  */
1369 unsigned long do_mmap(struct file *file, unsigned long addr,
1370                         unsigned long len, unsigned long prot,
1371                         unsigned long flags, unsigned long pgoff,
1372                         unsigned long *populate, struct list_head *uf)
1373 {
1374         struct mm_struct *mm = current->mm;
1375         vm_flags_t vm_flags;
1376         int pkey = 0;
1377
1378         *populate = 0;
1379
1380         if (!len)
1381                 return -EINVAL;
1382
1383         /*
1384          * Does the application expect PROT_READ to imply PROT_EXEC?
1385          *
1386          * (the exception is when the underlying filesystem is noexec
1387          *  mounted, in which case we dont add PROT_EXEC.)
1388          */
1389         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1390                 if (!(file && path_noexec(&file->f_path)))
1391                         prot |= PROT_EXEC;
1392
1393         /* force arch specific MAP_FIXED handling in get_unmapped_area */
1394         if (flags & MAP_FIXED_NOREPLACE)
1395                 flags |= MAP_FIXED;
1396
1397         if (!(flags & MAP_FIXED))
1398                 addr = round_hint_to_min(addr);
1399
1400         /* Careful about overflows.. */
1401         len = PAGE_ALIGN(len);
1402         if (!len)
1403                 return -ENOMEM;
1404
1405         /* offset overflow? */
1406         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1407                 return -EOVERFLOW;
1408
1409         /* Too many mappings? */
1410         if (mm->map_count > sysctl_max_map_count)
1411                 return -ENOMEM;
1412
1413         /* Obtain the address to map to. we verify (or select) it and ensure
1414          * that it represents a valid section of the address space.
1415          */
1416         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1417         if (IS_ERR_VALUE(addr))
1418                 return addr;
1419
1420         if (flags & MAP_FIXED_NOREPLACE) {
1421                 if (find_vma_intersection(mm, addr, addr + len))
1422                         return -EEXIST;
1423         }
1424
1425         if (prot == PROT_EXEC) {
1426                 pkey = execute_only_pkey(mm);
1427                 if (pkey < 0)
1428                         pkey = 0;
1429         }
1430
1431         /* Do simple checking here so the lower-level routines won't have
1432          * to. we assume access permissions have been handled by the open
1433          * of the memory object, so we don't do any here.
1434          */
1435         vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1436                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1437
1438         if (flags & MAP_LOCKED)
1439                 if (!can_do_mlock())
1440                         return -EPERM;
1441
1442         if (mlock_future_check(mm, vm_flags, len))
1443                 return -EAGAIN;
1444
1445         if (file) {
1446                 struct inode *inode = file_inode(file);
1447                 unsigned long flags_mask;
1448
1449                 if (!file_mmap_ok(file, inode, pgoff, len))
1450                         return -EOVERFLOW;
1451
1452                 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1453
1454                 switch (flags & MAP_TYPE) {
1455                 case MAP_SHARED:
1456                         /*
1457                          * Force use of MAP_SHARED_VALIDATE with non-legacy
1458                          * flags. E.g. MAP_SYNC is dangerous to use with
1459                          * MAP_SHARED as you don't know which consistency model
1460                          * you will get. We silently ignore unsupported flags
1461                          * with MAP_SHARED to preserve backward compatibility.
1462                          */
1463                         flags &= LEGACY_MAP_MASK;
1464                         fallthrough;
1465                 case MAP_SHARED_VALIDATE:
1466                         if (flags & ~flags_mask)
1467                                 return -EOPNOTSUPP;
1468                         if (prot & PROT_WRITE) {
1469                                 if (!(file->f_mode & FMODE_WRITE))
1470                                         return -EACCES;
1471                                 if (IS_SWAPFILE(file->f_mapping->host))
1472                                         return -ETXTBSY;
1473                         }
1474
1475                         /*
1476                          * Make sure we don't allow writing to an append-only
1477                          * file..
1478                          */
1479                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1480                                 return -EACCES;
1481
1482                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1483                         if (!(file->f_mode & FMODE_WRITE))
1484                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1485                         fallthrough;
1486                 case MAP_PRIVATE:
1487                         if (!(file->f_mode & FMODE_READ))
1488                                 return -EACCES;
1489                         if (path_noexec(&file->f_path)) {
1490                                 if (vm_flags & VM_EXEC)
1491                                         return -EPERM;
1492                                 vm_flags &= ~VM_MAYEXEC;
1493                         }
1494
1495                         if (!file->f_op->mmap)
1496                                 return -ENODEV;
1497                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1498                                 return -EINVAL;
1499                         break;
1500
1501                 default:
1502                         return -EINVAL;
1503                 }
1504         } else {
1505                 switch (flags & MAP_TYPE) {
1506                 case MAP_SHARED:
1507                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1508                                 return -EINVAL;
1509                         /*
1510                          * Ignore pgoff.
1511                          */
1512                         pgoff = 0;
1513                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1514                         break;
1515                 case MAP_PRIVATE:
1516                         /*
1517                          * Set pgoff according to addr for anon_vma.
1518                          */
1519                         pgoff = addr >> PAGE_SHIFT;
1520                         break;
1521                 default:
1522                         return -EINVAL;
1523                 }
1524         }
1525
1526         /*
1527          * Set 'VM_NORESERVE' if we should not account for the
1528          * memory use of this mapping.
1529          */
1530         if (flags & MAP_NORESERVE) {
1531                 /* We honor MAP_NORESERVE if allowed to overcommit */
1532                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1533                         vm_flags |= VM_NORESERVE;
1534
1535                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1536                 if (file && is_file_hugepages(file))
1537                         vm_flags |= VM_NORESERVE;
1538         }
1539
1540         addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1541         if (!IS_ERR_VALUE(addr) &&
1542             ((vm_flags & VM_LOCKED) ||
1543              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1544                 *populate = len;
1545         return addr;
1546 }
1547
1548 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1549                               unsigned long prot, unsigned long flags,
1550                               unsigned long fd, unsigned long pgoff)
1551 {
1552         struct file *file = NULL;
1553         unsigned long retval;
1554
1555         if (!(flags & MAP_ANONYMOUS)) {
1556                 audit_mmap_fd(fd, flags);
1557                 file = fget(fd);
1558                 if (!file)
1559                         return -EBADF;
1560                 if (is_file_hugepages(file)) {
1561                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1562                 } else if (unlikely(flags & MAP_HUGETLB)) {
1563                         retval = -EINVAL;
1564                         goto out_fput;
1565                 }
1566         } else if (flags & MAP_HUGETLB) {
1567                 struct hstate *hs;
1568
1569                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1570                 if (!hs)
1571                         return -EINVAL;
1572
1573                 len = ALIGN(len, huge_page_size(hs));
1574                 /*
1575                  * VM_NORESERVE is used because the reservations will be
1576                  * taken when vm_ops->mmap() is called
1577                  */
1578                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1579                                 VM_NORESERVE,
1580                                 HUGETLB_ANONHUGE_INODE,
1581                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1582                 if (IS_ERR(file))
1583                         return PTR_ERR(file);
1584         }
1585
1586         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1587 out_fput:
1588         if (file)
1589                 fput(file);
1590         return retval;
1591 }
1592
1593 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1594                 unsigned long, prot, unsigned long, flags,
1595                 unsigned long, fd, unsigned long, pgoff)
1596 {
1597         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1598 }
1599
1600 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1601 struct mmap_arg_struct {
1602         unsigned long addr;
1603         unsigned long len;
1604         unsigned long prot;
1605         unsigned long flags;
1606         unsigned long fd;
1607         unsigned long offset;
1608 };
1609
1610 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1611 {
1612         struct mmap_arg_struct a;
1613
1614         if (copy_from_user(&a, arg, sizeof(a)))
1615                 return -EFAULT;
1616         if (offset_in_page(a.offset))
1617                 return -EINVAL;
1618
1619         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1620                                a.offset >> PAGE_SHIFT);
1621 }
1622 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1623
1624 /*
1625  * Some shared mappings will want the pages marked read-only
1626  * to track write events. If so, we'll downgrade vm_page_prot
1627  * to the private version (using protection_map[] without the
1628  * VM_SHARED bit).
1629  */
1630 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1631 {
1632         vm_flags_t vm_flags = vma->vm_flags;
1633         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1634
1635         /* If it was private or non-writable, the write bit is already clear */
1636         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1637                 return 0;
1638
1639         /* The backer wishes to know when pages are first written to? */
1640         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1641                 return 1;
1642
1643         /* The open routine did something to the protections that pgprot_modify
1644          * won't preserve? */
1645         if (pgprot_val(vm_page_prot) !=
1646             pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1647                 return 0;
1648
1649         /* Do we need to track softdirty? */
1650         if (vma_soft_dirty_enabled(vma))
1651                 return 1;
1652
1653         /* Specialty mapping? */
1654         if (vm_flags & VM_PFNMAP)
1655                 return 0;
1656
1657         /* Can the mapping track the dirty pages? */
1658         return vma->vm_file && vma->vm_file->f_mapping &&
1659                 mapping_can_writeback(vma->vm_file->f_mapping);
1660 }
1661
1662 /*
1663  * We account for memory if it's a private writeable mapping,
1664  * not hugepages and VM_NORESERVE wasn't set.
1665  */
1666 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1667 {
1668         /*
1669          * hugetlb has its own accounting separate from the core VM
1670          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1671          */
1672         if (file && is_file_hugepages(file))
1673                 return 0;
1674
1675         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1676 }
1677
1678 unsigned long mmap_region(struct file *file, unsigned long addr,
1679                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1680                 struct list_head *uf)
1681 {
1682         struct mm_struct *mm = current->mm;
1683         struct vm_area_struct *vma, *prev, *merge;
1684         int error;
1685         struct rb_node **rb_link, *rb_parent;
1686         unsigned long charged = 0;
1687
1688         /* Check against address space limit. */
1689         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1690                 unsigned long nr_pages;
1691
1692                 /*
1693                  * MAP_FIXED may remove pages of mappings that intersects with
1694                  * requested mapping. Account for the pages it would unmap.
1695                  */
1696                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1697
1698                 if (!may_expand_vm(mm, vm_flags,
1699                                         (len >> PAGE_SHIFT) - nr_pages))
1700                         return -ENOMEM;
1701         }
1702
1703         /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1704         if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1705                 return -ENOMEM;
1706         /*
1707          * Private writable mapping: check memory availability
1708          */
1709         if (accountable_mapping(file, vm_flags)) {
1710                 charged = len >> PAGE_SHIFT;
1711                 if (security_vm_enough_memory_mm(mm, charged))
1712                         return -ENOMEM;
1713                 vm_flags |= VM_ACCOUNT;
1714         }
1715
1716         /*
1717          * Can we just expand an old mapping?
1718          */
1719         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1720                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
1721         if (vma)
1722                 goto out;
1723
1724         /*
1725          * Determine the object being mapped and call the appropriate
1726          * specific mapper. the address has already been validated, but
1727          * not unmapped, but the maps are removed from the list.
1728          */
1729         vma = vm_area_alloc(mm);
1730         if (!vma) {
1731                 error = -ENOMEM;
1732                 goto unacct_error;
1733         }
1734
1735         vma->vm_start = addr;
1736         vma->vm_end = addr + len;
1737         vma->vm_flags = vm_flags;
1738         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1739         vma->vm_pgoff = pgoff;
1740
1741         if (file) {
1742                 if (vm_flags & VM_SHARED) {
1743                         error = mapping_map_writable(file->f_mapping);
1744                         if (error)
1745                                 goto free_vma;
1746                 }
1747
1748                 vma->vm_file = get_file(file);
1749                 error = call_mmap(file, vma);
1750                 if (error)
1751                         goto unmap_and_free_vma;
1752
1753                 /* Can addr have changed??
1754                  *
1755                  * Answer: Yes, several device drivers can do it in their
1756                  *         f_op->mmap method. -DaveM
1757                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1758                  *      be updated for vma_link()
1759                  */
1760                 WARN_ON_ONCE(addr != vma->vm_start);
1761
1762                 addr = vma->vm_start;
1763
1764                 /* If vm_flags changed after call_mmap(), we should try merge vma again
1765                  * as we may succeed this time.
1766                  */
1767                 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1768                         merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1769                                 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
1770                         if (merge) {
1771                                 /* ->mmap() can change vma->vm_file and fput the original file. So
1772                                  * fput the vma->vm_file here or we would add an extra fput for file
1773                                  * and cause general protection fault ultimately.
1774                                  */
1775                                 fput(vma->vm_file);
1776                                 vm_area_free(vma);
1777                                 vma = merge;
1778                                 /* Update vm_flags to pick up the change. */
1779                                 vm_flags = vma->vm_flags;
1780                                 goto unmap_writable;
1781                         }
1782                 }
1783
1784                 vm_flags = vma->vm_flags;
1785         } else if (vm_flags & VM_SHARED) {
1786                 error = shmem_zero_setup(vma);
1787                 if (error)
1788                         goto free_vma;
1789         } else {
1790                 vma_set_anonymous(vma);
1791         }
1792
1793         /* Allow architectures to sanity-check the vm_flags */
1794         if (!arch_validate_flags(vma->vm_flags)) {
1795                 error = -EINVAL;
1796                 if (file)
1797                         goto unmap_and_free_vma;
1798                 else
1799                         goto free_vma;
1800         }
1801
1802         vma_link(mm, vma, prev, rb_link, rb_parent);
1803
1804         /*
1805          * vma_merge() calls khugepaged_enter_vma() either, the below
1806          * call covers the non-merge case.
1807          */
1808         khugepaged_enter_vma(vma, vma->vm_flags);
1809
1810         /* Once vma denies write, undo our temporary denial count */
1811 unmap_writable:
1812         if (file && vm_flags & VM_SHARED)
1813                 mapping_unmap_writable(file->f_mapping);
1814         file = vma->vm_file;
1815 out:
1816         perf_event_mmap(vma);
1817
1818         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1819         if (vm_flags & VM_LOCKED) {
1820                 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1821                                         is_vm_hugetlb_page(vma) ||
1822                                         vma == get_gate_vma(current->mm))
1823                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1824                 else
1825                         mm->locked_vm += (len >> PAGE_SHIFT);
1826         }
1827
1828         if (file)
1829                 uprobe_mmap(vma);
1830
1831         /*
1832          * New (or expanded) vma always get soft dirty status.
1833          * Otherwise user-space soft-dirty page tracker won't
1834          * be able to distinguish situation when vma area unmapped,
1835          * then new mapped in-place (which must be aimed as
1836          * a completely new data area).
1837          */
1838         vma->vm_flags |= VM_SOFTDIRTY;
1839
1840         vma_set_page_prot(vma);
1841
1842         return addr;
1843
1844 unmap_and_free_vma:
1845         fput(vma->vm_file);
1846         vma->vm_file = NULL;
1847
1848         /* Undo any partial mapping done by a device driver. */
1849         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1850         if (vm_flags & VM_SHARED)
1851                 mapping_unmap_writable(file->f_mapping);
1852 free_vma:
1853         vm_area_free(vma);
1854 unacct_error:
1855         if (charged)
1856                 vm_unacct_memory(charged);
1857         return error;
1858 }
1859
1860 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1861 {
1862         /*
1863          * We implement the search by looking for an rbtree node that
1864          * immediately follows a suitable gap. That is,
1865          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1866          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1867          * - gap_end - gap_start >= length
1868          */
1869
1870         struct mm_struct *mm = current->mm;
1871         struct vm_area_struct *vma;
1872         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1873
1874         /* Adjust search length to account for worst case alignment overhead */
1875         length = info->length + info->align_mask;
1876         if (length < info->length)
1877                 return -ENOMEM;
1878
1879         /* Adjust search limits by the desired length */
1880         if (info->high_limit < length)
1881                 return -ENOMEM;
1882         high_limit = info->high_limit - length;
1883
1884         if (info->low_limit > high_limit)
1885                 return -ENOMEM;
1886         low_limit = info->low_limit + length;
1887
1888         /* Check if rbtree root looks promising */
1889         if (RB_EMPTY_ROOT(&mm->mm_rb))
1890                 goto check_highest;
1891         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1892         if (vma->rb_subtree_gap < length)
1893                 goto check_highest;
1894
1895         while (true) {
1896                 /* Visit left subtree if it looks promising */
1897                 gap_end = vm_start_gap(vma);
1898                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1899                         struct vm_area_struct *left =
1900                                 rb_entry(vma->vm_rb.rb_left,
1901                                          struct vm_area_struct, vm_rb);
1902                         if (left->rb_subtree_gap >= length) {
1903                                 vma = left;
1904                                 continue;
1905                         }
1906                 }
1907
1908                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1909 check_current:
1910                 /* Check if current node has a suitable gap */
1911                 if (gap_start > high_limit)
1912                         return -ENOMEM;
1913                 if (gap_end >= low_limit &&
1914                     gap_end > gap_start && gap_end - gap_start >= length)
1915                         goto found;
1916
1917                 /* Visit right subtree if it looks promising */
1918                 if (vma->vm_rb.rb_right) {
1919                         struct vm_area_struct *right =
1920                                 rb_entry(vma->vm_rb.rb_right,
1921                                          struct vm_area_struct, vm_rb);
1922                         if (right->rb_subtree_gap >= length) {
1923                                 vma = right;
1924                                 continue;
1925                         }
1926                 }
1927
1928                 /* Go back up the rbtree to find next candidate node */
1929                 while (true) {
1930                         struct rb_node *prev = &vma->vm_rb;
1931                         if (!rb_parent(prev))
1932                                 goto check_highest;
1933                         vma = rb_entry(rb_parent(prev),
1934                                        struct vm_area_struct, vm_rb);
1935                         if (prev == vma->vm_rb.rb_left) {
1936                                 gap_start = vm_end_gap(vma->vm_prev);
1937                                 gap_end = vm_start_gap(vma);
1938                                 goto check_current;
1939                         }
1940                 }
1941         }
1942
1943 check_highest:
1944         /* Check highest gap, which does not precede any rbtree node */
1945         gap_start = mm->highest_vm_end;
1946         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1947         if (gap_start > high_limit)
1948                 return -ENOMEM;
1949
1950 found:
1951         /* We found a suitable gap. Clip it with the original low_limit. */
1952         if (gap_start < info->low_limit)
1953                 gap_start = info->low_limit;
1954
1955         /* Adjust gap address to the desired alignment */
1956         gap_start += (info->align_offset - gap_start) & info->align_mask;
1957
1958         VM_BUG_ON(gap_start + info->length > info->high_limit);
1959         VM_BUG_ON(gap_start + info->length > gap_end);
1960         return gap_start;
1961 }
1962
1963 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1964 {
1965         struct mm_struct *mm = current->mm;
1966         struct vm_area_struct *vma;
1967         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1968
1969         /* Adjust search length to account for worst case alignment overhead */
1970         length = info->length + info->align_mask;
1971         if (length < info->length)
1972                 return -ENOMEM;
1973
1974         /*
1975          * Adjust search limits by the desired length.
1976          * See implementation comment at top of unmapped_area().
1977          */
1978         gap_end = info->high_limit;
1979         if (gap_end < length)
1980                 return -ENOMEM;
1981         high_limit = gap_end - length;
1982
1983         if (info->low_limit > high_limit)
1984                 return -ENOMEM;
1985         low_limit = info->low_limit + length;
1986
1987         /* Check highest gap, which does not precede any rbtree node */
1988         gap_start = mm->highest_vm_end;
1989         if (gap_start <= high_limit)
1990                 goto found_highest;
1991
1992         /* Check if rbtree root looks promising */
1993         if (RB_EMPTY_ROOT(&mm->mm_rb))
1994                 return -ENOMEM;
1995         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1996         if (vma->rb_subtree_gap < length)
1997                 return -ENOMEM;
1998
1999         while (true) {
2000                 /* Visit right subtree if it looks promising */
2001                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2002                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2003                         struct vm_area_struct *right =
2004                                 rb_entry(vma->vm_rb.rb_right,
2005                                          struct vm_area_struct, vm_rb);
2006                         if (right->rb_subtree_gap >= length) {
2007                                 vma = right;
2008                                 continue;
2009                         }
2010                 }
2011
2012 check_current:
2013                 /* Check if current node has a suitable gap */
2014                 gap_end = vm_start_gap(vma);
2015                 if (gap_end < low_limit)
2016                         return -ENOMEM;
2017                 if (gap_start <= high_limit &&
2018                     gap_end > gap_start && gap_end - gap_start >= length)
2019                         goto found;
2020
2021                 /* Visit left subtree if it looks promising */
2022                 if (vma->vm_rb.rb_left) {
2023                         struct vm_area_struct *left =
2024                                 rb_entry(vma->vm_rb.rb_left,
2025                                          struct vm_area_struct, vm_rb);
2026                         if (left->rb_subtree_gap >= length) {
2027                                 vma = left;
2028                                 continue;
2029                         }
2030                 }
2031
2032                 /* Go back up the rbtree to find next candidate node */
2033                 while (true) {
2034                         struct rb_node *prev = &vma->vm_rb;
2035                         if (!rb_parent(prev))
2036                                 return -ENOMEM;
2037                         vma = rb_entry(rb_parent(prev),
2038                                        struct vm_area_struct, vm_rb);
2039                         if (prev == vma->vm_rb.rb_right) {
2040                                 gap_start = vma->vm_prev ?
2041                                         vm_end_gap(vma->vm_prev) : 0;
2042                                 goto check_current;
2043                         }
2044                 }
2045         }
2046
2047 found:
2048         /* We found a suitable gap. Clip it with the original high_limit. */
2049         if (gap_end > info->high_limit)
2050                 gap_end = info->high_limit;
2051
2052 found_highest:
2053         /* Compute highest gap address at the desired alignment */
2054         gap_end -= info->length;
2055         gap_end -= (gap_end - info->align_offset) & info->align_mask;
2056
2057         VM_BUG_ON(gap_end < info->low_limit);
2058         VM_BUG_ON(gap_end < gap_start);
2059         return gap_end;
2060 }
2061
2062 /*
2063  * Search for an unmapped address range.
2064  *
2065  * We are looking for a range that:
2066  * - does not intersect with any VMA;
2067  * - is contained within the [low_limit, high_limit) interval;
2068  * - is at least the desired size.
2069  * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2070  */
2071 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2072 {
2073         unsigned long addr;
2074
2075         if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2076                 addr = unmapped_area_topdown(info);
2077         else
2078                 addr = unmapped_area(info);
2079
2080         trace_vm_unmapped_area(addr, info);
2081         return addr;
2082 }
2083
2084 /* Get an address range which is currently unmapped.
2085  * For shmat() with addr=0.
2086  *
2087  * Ugly calling convention alert:
2088  * Return value with the low bits set means error value,
2089  * ie
2090  *      if (ret & ~PAGE_MASK)
2091  *              error = ret;
2092  *
2093  * This function "knows" that -ENOMEM has the bits set.
2094  */
2095 unsigned long
2096 generic_get_unmapped_area(struct file *filp, unsigned long addr,
2097                           unsigned long len, unsigned long pgoff,
2098                           unsigned long flags)
2099 {
2100         struct mm_struct *mm = current->mm;
2101         struct vm_area_struct *vma, *prev;
2102         struct vm_unmapped_area_info info;
2103         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
2104
2105         if (len > mmap_end - mmap_min_addr)
2106                 return -ENOMEM;
2107
2108         if (flags & MAP_FIXED)
2109                 return addr;
2110
2111         if (addr) {
2112                 addr = PAGE_ALIGN(addr);
2113                 vma = find_vma_prev(mm, addr, &prev);
2114                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2115                     (!vma || addr + len <= vm_start_gap(vma)) &&
2116                     (!prev || addr >= vm_end_gap(prev)))
2117                         return addr;
2118         }
2119
2120         info.flags = 0;
2121         info.length = len;
2122         info.low_limit = mm->mmap_base;
2123         info.high_limit = mmap_end;
2124         info.align_mask = 0;
2125         info.align_offset = 0;
2126         return vm_unmapped_area(&info);
2127 }
2128
2129 #ifndef HAVE_ARCH_UNMAPPED_AREA
2130 unsigned long
2131 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2132                        unsigned long len, unsigned long pgoff,
2133                        unsigned long flags)
2134 {
2135         return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
2136 }
2137 #endif
2138
2139 /*
2140  * This mmap-allocator allocates new areas top-down from below the
2141  * stack's low limit (the base):
2142  */
2143 unsigned long
2144 generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2145                                   unsigned long len, unsigned long pgoff,
2146                                   unsigned long flags)
2147 {
2148         struct vm_area_struct *vma, *prev;
2149         struct mm_struct *mm = current->mm;
2150         struct vm_unmapped_area_info info;
2151         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
2152
2153         /* requested length too big for entire address space */
2154         if (len > mmap_end - mmap_min_addr)
2155                 return -ENOMEM;
2156
2157         if (flags & MAP_FIXED)
2158                 return addr;
2159
2160         /* requesting a specific address */
2161         if (addr) {
2162                 addr = PAGE_ALIGN(addr);
2163                 vma = find_vma_prev(mm, addr, &prev);
2164                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2165                                 (!vma || addr + len <= vm_start_gap(vma)) &&
2166                                 (!prev || addr >= vm_end_gap(prev)))
2167                         return addr;
2168         }
2169
2170         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2171         info.length = len;
2172         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2173         info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2174         info.align_mask = 0;
2175         info.align_offset = 0;
2176         addr = vm_unmapped_area(&info);
2177
2178         /*
2179          * A failed mmap() very likely causes application failure,
2180          * so fall back to the bottom-up function here. This scenario
2181          * can happen with large stack limits and large mmap()
2182          * allocations.
2183          */
2184         if (offset_in_page(addr)) {
2185                 VM_BUG_ON(addr != -ENOMEM);
2186                 info.flags = 0;
2187                 info.low_limit = TASK_UNMAPPED_BASE;
2188                 info.high_limit = mmap_end;
2189                 addr = vm_unmapped_area(&info);
2190         }
2191
2192         return addr;
2193 }
2194
2195 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2196 unsigned long
2197 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2198                                unsigned long len, unsigned long pgoff,
2199                                unsigned long flags)
2200 {
2201         return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
2202 }
2203 #endif
2204
2205 unsigned long
2206 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2207                 unsigned long pgoff, unsigned long flags)
2208 {
2209         unsigned long (*get_area)(struct file *, unsigned long,
2210                                   unsigned long, unsigned long, unsigned long);
2211
2212         unsigned long error = arch_mmap_check(addr, len, flags);
2213         if (error)
2214                 return error;
2215
2216         /* Careful about overflows.. */
2217         if (len > TASK_SIZE)
2218                 return -ENOMEM;
2219
2220         get_area = current->mm->get_unmapped_area;
2221         if (file) {
2222                 if (file->f_op->get_unmapped_area)
2223                         get_area = file->f_op->get_unmapped_area;
2224         } else if (flags & MAP_SHARED) {
2225                 /*
2226                  * mmap_region() will call shmem_zero_setup() to create a file,
2227                  * so use shmem's get_unmapped_area in case it can be huge.
2228                  * do_mmap() will clear pgoff, so match alignment.
2229                  */
2230                 pgoff = 0;
2231                 get_area = shmem_get_unmapped_area;
2232         }
2233
2234         addr = get_area(file, addr, len, pgoff, flags);
2235         if (IS_ERR_VALUE(addr))
2236                 return addr;
2237
2238         if (addr > TASK_SIZE - len)
2239                 return -ENOMEM;
2240         if (offset_in_page(addr))
2241                 return -EINVAL;
2242
2243         error = security_mmap_addr(addr);
2244         return error ? error : addr;
2245 }
2246
2247 EXPORT_SYMBOL(get_unmapped_area);
2248
2249 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2250 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2251 {
2252         struct rb_node *rb_node;
2253         struct vm_area_struct *vma;
2254
2255         mmap_assert_locked(mm);
2256         /* Check the cache first. */
2257         vma = vmacache_find(mm, addr);
2258         if (likely(vma))
2259                 return vma;
2260
2261         rb_node = mm->mm_rb.rb_node;
2262
2263         while (rb_node) {
2264                 struct vm_area_struct *tmp;
2265
2266                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2267
2268                 if (tmp->vm_end > addr) {
2269                         vma = tmp;
2270                         if (tmp->vm_start <= addr)
2271                                 break;
2272                         rb_node = rb_node->rb_left;
2273                 } else
2274                         rb_node = rb_node->rb_right;
2275         }
2276
2277         if (vma)
2278                 vmacache_update(addr, vma);
2279         return vma;
2280 }
2281
2282 EXPORT_SYMBOL(find_vma);
2283
2284 /*
2285  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2286  */
2287 struct vm_area_struct *
2288 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2289                         struct vm_area_struct **pprev)
2290 {
2291         struct vm_area_struct *vma;
2292
2293         vma = find_vma(mm, addr);
2294         if (vma) {
2295                 *pprev = vma->vm_prev;
2296         } else {
2297                 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2298
2299                 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2300         }
2301         return vma;
2302 }
2303
2304 /*
2305  * Verify that the stack growth is acceptable and
2306  * update accounting. This is shared with both the
2307  * grow-up and grow-down cases.
2308  */
2309 static int acct_stack_growth(struct vm_area_struct *vma,
2310                              unsigned long size, unsigned long grow)
2311 {
2312         struct mm_struct *mm = vma->vm_mm;
2313         unsigned long new_start;
2314
2315         /* address space limit tests */
2316         if (!may_expand_vm(mm, vma->vm_flags, grow))
2317                 return -ENOMEM;
2318
2319         /* Stack limit test */
2320         if (size > rlimit(RLIMIT_STACK))
2321                 return -ENOMEM;
2322
2323         /* mlock limit tests */
2324         if (mlock_future_check(mm, vma->vm_flags, grow << PAGE_SHIFT))
2325                 return -ENOMEM;
2326
2327         /* Check to ensure the stack will not grow into a hugetlb-only region */
2328         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2329                         vma->vm_end - size;
2330         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2331                 return -EFAULT;
2332
2333         /*
2334          * Overcommit..  This must be the final test, as it will
2335          * update security statistics.
2336          */
2337         if (security_vm_enough_memory_mm(mm, grow))
2338                 return -ENOMEM;
2339
2340         return 0;
2341 }
2342
2343 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2344 /*
2345  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2346  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2347  */
2348 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2349 {
2350         struct mm_struct *mm = vma->vm_mm;
2351         struct vm_area_struct *next;
2352         unsigned long gap_addr;
2353         int error = 0;
2354
2355         if (!(vma->vm_flags & VM_GROWSUP))
2356                 return -EFAULT;
2357
2358         /* Guard against exceeding limits of the address space. */
2359         address &= PAGE_MASK;
2360         if (address >= (TASK_SIZE & PAGE_MASK))
2361                 return -ENOMEM;
2362         address += PAGE_SIZE;
2363
2364         /* Enforce stack_guard_gap */
2365         gap_addr = address + stack_guard_gap;
2366
2367         /* Guard against overflow */
2368         if (gap_addr < address || gap_addr > TASK_SIZE)
2369                 gap_addr = TASK_SIZE;
2370
2371         next = vma->vm_next;
2372         if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2373                 if (!(next->vm_flags & VM_GROWSUP))
2374                         return -ENOMEM;
2375                 /* Check that both stack segments have the same anon_vma? */
2376         }
2377
2378         /* We must make sure the anon_vma is allocated. */
2379         if (unlikely(anon_vma_prepare(vma)))
2380                 return -ENOMEM;
2381
2382         /*
2383          * vma->vm_start/vm_end cannot change under us because the caller
2384          * is required to hold the mmap_lock in read mode.  We need the
2385          * anon_vma lock to serialize against concurrent expand_stacks.
2386          */
2387         anon_vma_lock_write(vma->anon_vma);
2388
2389         /* Somebody else might have raced and expanded it already */
2390         if (address > vma->vm_end) {
2391                 unsigned long size, grow;
2392
2393                 size = address - vma->vm_start;
2394                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2395
2396                 error = -ENOMEM;
2397                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2398                         error = acct_stack_growth(vma, size, grow);
2399                         if (!error) {
2400                                 /*
2401                                  * vma_gap_update() doesn't support concurrent
2402                                  * updates, but we only hold a shared mmap_lock
2403                                  * lock here, so we need to protect against
2404                                  * concurrent vma expansions.
2405                                  * anon_vma_lock_write() doesn't help here, as
2406                                  * we don't guarantee that all growable vmas
2407                                  * in a mm share the same root anon vma.
2408                                  * So, we reuse mm->page_table_lock to guard
2409                                  * against concurrent vma expansions.
2410                                  */
2411                                 spin_lock(&mm->page_table_lock);
2412                                 if (vma->vm_flags & VM_LOCKED)
2413                                         mm->locked_vm += grow;
2414                                 vm_stat_account(mm, vma->vm_flags, grow);
2415                                 anon_vma_interval_tree_pre_update_vma(vma);
2416                                 vma->vm_end = address;
2417                                 anon_vma_interval_tree_post_update_vma(vma);
2418                                 if (vma->vm_next)
2419                                         vma_gap_update(vma->vm_next);
2420                                 else
2421                                         mm->highest_vm_end = vm_end_gap(vma);
2422                                 spin_unlock(&mm->page_table_lock);
2423
2424                                 perf_event_mmap(vma);
2425                         }
2426                 }
2427         }
2428         anon_vma_unlock_write(vma->anon_vma);
2429         khugepaged_enter_vma(vma, vma->vm_flags);
2430         validate_mm(mm);
2431         return error;
2432 }
2433 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2434
2435 /*
2436  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2437  */
2438 int expand_downwards(struct vm_area_struct *vma,
2439                                    unsigned long address)
2440 {
2441         struct mm_struct *mm = vma->vm_mm;
2442         struct vm_area_struct *prev;
2443         int error = 0;
2444
2445         address &= PAGE_MASK;
2446         if (address < mmap_min_addr)
2447                 return -EPERM;
2448
2449         /* Enforce stack_guard_gap */
2450         prev = vma->vm_prev;
2451         /* Check that both stack segments have the same anon_vma? */
2452         if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2453                         vma_is_accessible(prev)) {
2454                 if (address - prev->vm_end < stack_guard_gap)
2455                         return -ENOMEM;
2456         }
2457
2458         /* We must make sure the anon_vma is allocated. */
2459         if (unlikely(anon_vma_prepare(vma)))
2460                 return -ENOMEM;
2461
2462         /*
2463          * vma->vm_start/vm_end cannot change under us because the caller
2464          * is required to hold the mmap_lock in read mode.  We need the
2465          * anon_vma lock to serialize against concurrent expand_stacks.
2466          */
2467         anon_vma_lock_write(vma->anon_vma);
2468
2469         /* Somebody else might have raced and expanded it already */
2470         if (address < vma->vm_start) {
2471                 unsigned long size, grow;
2472
2473                 size = vma->vm_end - address;
2474                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2475
2476                 error = -ENOMEM;
2477                 if (grow <= vma->vm_pgoff) {
2478                         error = acct_stack_growth(vma, size, grow);
2479                         if (!error) {
2480                                 /*
2481                                  * vma_gap_update() doesn't support concurrent
2482                                  * updates, but we only hold a shared mmap_lock
2483                                  * lock here, so we need to protect against
2484                                  * concurrent vma expansions.
2485                                  * anon_vma_lock_write() doesn't help here, as
2486                                  * we don't guarantee that all growable vmas
2487                                  * in a mm share the same root anon vma.
2488                                  * So, we reuse mm->page_table_lock to guard
2489                                  * against concurrent vma expansions.
2490                                  */
2491                                 spin_lock(&mm->page_table_lock);
2492                                 if (vma->vm_flags & VM_LOCKED)
2493                                         mm->locked_vm += grow;
2494                                 vm_stat_account(mm, vma->vm_flags, grow);
2495                                 anon_vma_interval_tree_pre_update_vma(vma);
2496                                 vma->vm_start = address;
2497                                 vma->vm_pgoff -= grow;
2498                                 anon_vma_interval_tree_post_update_vma(vma);
2499                                 vma_gap_update(vma);
2500                                 spin_unlock(&mm->page_table_lock);
2501
2502                                 perf_event_mmap(vma);
2503                         }
2504                 }
2505         }
2506         anon_vma_unlock_write(vma->anon_vma);
2507         khugepaged_enter_vma(vma, vma->vm_flags);
2508         validate_mm(mm);
2509         return error;
2510 }
2511
2512 /* enforced gap between the expanding stack and other mappings. */
2513 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2514
2515 static int __init cmdline_parse_stack_guard_gap(char *p)
2516 {
2517         unsigned long val;
2518         char *endptr;
2519
2520         val = simple_strtoul(p, &endptr, 10);
2521         if (!*endptr)
2522                 stack_guard_gap = val << PAGE_SHIFT;
2523
2524         return 1;
2525 }
2526 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2527
2528 #ifdef CONFIG_STACK_GROWSUP
2529 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2530 {
2531         return expand_upwards(vma, address);
2532 }
2533
2534 struct vm_area_struct *
2535 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2536 {
2537         struct vm_area_struct *vma, *prev;
2538
2539         addr &= PAGE_MASK;
2540         vma = find_vma_prev(mm, addr, &prev);
2541         if (vma && (vma->vm_start <= addr))
2542                 return vma;
2543         if (!prev || expand_stack(prev, addr))
2544                 return NULL;
2545         if (prev->vm_flags & VM_LOCKED)
2546                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2547         return prev;
2548 }
2549 #else
2550 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2551 {
2552         return expand_downwards(vma, address);
2553 }
2554
2555 struct vm_area_struct *
2556 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2557 {
2558         struct vm_area_struct *vma;
2559         unsigned long start;
2560
2561         addr &= PAGE_MASK;
2562         vma = find_vma(mm, addr);
2563         if (!vma)
2564                 return NULL;
2565         if (vma->vm_start <= addr)
2566                 return vma;
2567         if (!(vma->vm_flags & VM_GROWSDOWN))
2568                 return NULL;
2569         start = vma->vm_start;
2570         if (expand_stack(vma, addr))
2571                 return NULL;
2572         if (vma->vm_flags & VM_LOCKED)
2573                 populate_vma_page_range(vma, addr, start, NULL);
2574         return vma;
2575 }
2576 #endif
2577
2578 EXPORT_SYMBOL_GPL(find_extend_vma);
2579
2580 /*
2581  * Ok - we have the memory areas we should free on the vma list,
2582  * so release them, and do the vma updates.
2583  *
2584  * Called with the mm semaphore held.
2585  */
2586 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2587 {
2588         unsigned long nr_accounted = 0;
2589
2590         /* Update high watermark before we lower total_vm */
2591         update_hiwater_vm(mm);
2592         do {
2593                 long nrpages = vma_pages(vma);
2594
2595                 if (vma->vm_flags & VM_ACCOUNT)
2596                         nr_accounted += nrpages;
2597                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2598                 vma = remove_vma(vma);
2599         } while (vma);
2600         vm_unacct_memory(nr_accounted);
2601         validate_mm(mm);
2602 }
2603
2604 /*
2605  * Get rid of page table information in the indicated region.
2606  *
2607  * Called with the mm semaphore held.
2608  */
2609 static void unmap_region(struct mm_struct *mm,
2610                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2611                 unsigned long start, unsigned long end)
2612 {
2613         struct vm_area_struct *next = vma_next(mm, prev);
2614         struct mmu_gather tlb;
2615
2616         lru_add_drain();
2617         tlb_gather_mmu(&tlb, mm);
2618         update_hiwater_rss(mm);
2619         unmap_vmas(&tlb, vma, start, end);
2620         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2621                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2622         tlb_finish_mmu(&tlb);
2623 }
2624
2625 /*
2626  * Create a list of vma's touched by the unmap, removing them from the mm's
2627  * vma list as we go..
2628  */
2629 static bool
2630 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2631         struct vm_area_struct *prev, unsigned long end)
2632 {
2633         struct vm_area_struct **insertion_point;
2634         struct vm_area_struct *tail_vma = NULL;
2635
2636         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2637         vma->vm_prev = NULL;
2638         do {
2639                 vma_rb_erase(vma, &mm->mm_rb);
2640                 if (vma->vm_flags & VM_LOCKED)
2641                         mm->locked_vm -= vma_pages(vma);
2642                 mm->map_count--;
2643                 tail_vma = vma;
2644                 vma = vma->vm_next;
2645         } while (vma && vma->vm_start < end);
2646         *insertion_point = vma;
2647         if (vma) {
2648                 vma->vm_prev = prev;
2649                 vma_gap_update(vma);
2650         } else
2651                 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2652         tail_vma->vm_next = NULL;
2653
2654         /* Kill the cache */
2655         vmacache_invalidate(mm);
2656
2657         /*
2658          * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2659          * VM_GROWSUP VMA. Such VMAs can change their size under
2660          * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2661          */
2662         if (vma && (vma->vm_flags & VM_GROWSDOWN))
2663                 return false;
2664         if (prev && (prev->vm_flags & VM_GROWSUP))
2665                 return false;
2666         return true;
2667 }
2668
2669 /*
2670  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2671  * has already been checked or doesn't make sense to fail.
2672  */
2673 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2674                 unsigned long addr, int new_below)
2675 {
2676         struct vm_area_struct *new;
2677         int err;
2678
2679         if (vma->vm_ops && vma->vm_ops->may_split) {
2680                 err = vma->vm_ops->may_split(vma, addr);
2681                 if (err)
2682                         return err;
2683         }
2684
2685         new = vm_area_dup(vma);
2686         if (!new)
2687                 return -ENOMEM;
2688
2689         if (new_below)
2690                 new->vm_end = addr;
2691         else {
2692                 new->vm_start = addr;
2693                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2694         }
2695
2696         err = vma_dup_policy(vma, new);
2697         if (err)
2698                 goto out_free_vma;
2699
2700         err = anon_vma_clone(new, vma);
2701         if (err)
2702                 goto out_free_mpol;
2703
2704         if (new->vm_file)
2705                 get_file(new->vm_file);
2706
2707         if (new->vm_ops && new->vm_ops->open)
2708                 new->vm_ops->open(new);
2709
2710         if (new_below)
2711                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2712                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2713         else
2714                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2715
2716         /* Success. */
2717         if (!err)
2718                 return 0;
2719
2720         /* Clean everything up if vma_adjust failed. */
2721         if (new->vm_ops && new->vm_ops->close)
2722                 new->vm_ops->close(new);
2723         if (new->vm_file)
2724                 fput(new->vm_file);
2725         unlink_anon_vmas(new);
2726  out_free_mpol:
2727         mpol_put(vma_policy(new));
2728  out_free_vma:
2729         vm_area_free(new);
2730         return err;
2731 }
2732
2733 /*
2734  * Split a vma into two pieces at address 'addr', a new vma is allocated
2735  * either for the first part or the tail.
2736  */
2737 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2738               unsigned long addr, int new_below)
2739 {
2740         if (mm->map_count >= sysctl_max_map_count)
2741                 return -ENOMEM;
2742
2743         return __split_vma(mm, vma, addr, new_below);
2744 }
2745
2746 /* Munmap is split into 2 main parts -- this part which finds
2747  * what needs doing, and the areas themselves, which do the
2748  * work.  This now handles partial unmappings.
2749  * Jeremy Fitzhardinge <jeremy@goop.org>
2750  */
2751 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2752                 struct list_head *uf, bool downgrade)
2753 {
2754         unsigned long end;
2755         struct vm_area_struct *vma, *prev, *last;
2756
2757         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2758                 return -EINVAL;
2759
2760         len = PAGE_ALIGN(len);
2761         end = start + len;
2762         if (len == 0)
2763                 return -EINVAL;
2764
2765         /*
2766          * arch_unmap() might do unmaps itself.  It must be called
2767          * and finish any rbtree manipulation before this code
2768          * runs and also starts to manipulate the rbtree.
2769          */
2770         arch_unmap(mm, start, end);
2771
2772         /* Find the first overlapping VMA where start < vma->vm_end */
2773         vma = find_vma_intersection(mm, start, end);
2774         if (!vma)
2775                 return 0;
2776         prev = vma->vm_prev;
2777
2778         /*
2779          * If we need to split any vma, do it now to save pain later.
2780          *
2781          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2782          * unmapped vm_area_struct will remain in use: so lower split_vma
2783          * places tmp vma above, and higher split_vma places tmp vma below.
2784          */
2785         if (start > vma->vm_start) {
2786                 int error;
2787
2788                 /*
2789                  * Make sure that map_count on return from munmap() will
2790                  * not exceed its limit; but let map_count go just above
2791                  * its limit temporarily, to help free resources as expected.
2792                  */
2793                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2794                         return -ENOMEM;
2795
2796                 error = __split_vma(mm, vma, start, 0);
2797                 if (error)
2798                         return error;
2799                 prev = vma;
2800         }
2801
2802         /* Does it split the last one? */
2803         last = find_vma(mm, end);
2804         if (last && end > last->vm_start) {
2805                 int error = __split_vma(mm, last, end, 1);
2806                 if (error)
2807                         return error;
2808         }
2809         vma = vma_next(mm, prev);
2810
2811         if (unlikely(uf)) {
2812                 /*
2813                  * If userfaultfd_unmap_prep returns an error the vmas
2814                  * will remain split, but userland will get a
2815                  * highly unexpected error anyway. This is no
2816                  * different than the case where the first of the two
2817                  * __split_vma fails, but we don't undo the first
2818                  * split, despite we could. This is unlikely enough
2819                  * failure that it's not worth optimizing it for.
2820                  */
2821                 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2822                 if (error)
2823                         return error;
2824         }
2825
2826         /* Detach vmas from rbtree */
2827         if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2828                 downgrade = false;
2829
2830         if (downgrade)
2831                 mmap_write_downgrade(mm);
2832
2833         unmap_region(mm, vma, prev, start, end);
2834
2835         /* Fix up all other VM information */
2836         remove_vma_list(mm, vma);
2837
2838         return downgrade ? 1 : 0;
2839 }
2840
2841 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2842               struct list_head *uf)
2843 {
2844         return __do_munmap(mm, start, len, uf, false);
2845 }
2846
2847 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2848 {
2849         int ret;
2850         struct mm_struct *mm = current->mm;
2851         LIST_HEAD(uf);
2852
2853         if (mmap_write_lock_killable(mm))
2854                 return -EINTR;
2855
2856         ret = __do_munmap(mm, start, len, &uf, downgrade);
2857         /*
2858          * Returning 1 indicates mmap_lock is downgraded.
2859          * But 1 is not legal return value of vm_munmap() and munmap(), reset
2860          * it to 0 before return.
2861          */
2862         if (ret == 1) {
2863                 mmap_read_unlock(mm);
2864                 ret = 0;
2865         } else
2866                 mmap_write_unlock(mm);
2867
2868         userfaultfd_unmap_complete(mm, &uf);
2869         return ret;
2870 }
2871
2872 int vm_munmap(unsigned long start, size_t len)
2873 {
2874         return __vm_munmap(start, len, false);
2875 }
2876 EXPORT_SYMBOL(vm_munmap);
2877
2878 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2879 {
2880         addr = untagged_addr(addr);
2881         return __vm_munmap(addr, len, true);
2882 }
2883
2884
2885 /*
2886  * Emulation of deprecated remap_file_pages() syscall.
2887  */
2888 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2889                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2890 {
2891
2892         struct mm_struct *mm = current->mm;
2893         struct vm_area_struct *vma;
2894         unsigned long populate = 0;
2895         unsigned long ret = -EINVAL;
2896         struct file *file;
2897
2898         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
2899                      current->comm, current->pid);
2900
2901         if (prot)
2902                 return ret;
2903         start = start & PAGE_MASK;
2904         size = size & PAGE_MASK;
2905
2906         if (start + size <= start)
2907                 return ret;
2908
2909         /* Does pgoff wrap? */
2910         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2911                 return ret;
2912
2913         if (mmap_write_lock_killable(mm))
2914                 return -EINTR;
2915
2916         vma = vma_lookup(mm, start);
2917
2918         if (!vma || !(vma->vm_flags & VM_SHARED))
2919                 goto out;
2920
2921         if (start + size > vma->vm_end) {
2922                 struct vm_area_struct *next;
2923
2924                 for (next = vma->vm_next; next; next = next->vm_next) {
2925                         /* hole between vmas ? */
2926                         if (next->vm_start != next->vm_prev->vm_end)
2927                                 goto out;
2928
2929                         if (next->vm_file != vma->vm_file)
2930                                 goto out;
2931
2932                         if (next->vm_flags != vma->vm_flags)
2933                                 goto out;
2934
2935                         if (start + size <= next->vm_end)
2936                                 break;
2937                 }
2938
2939                 if (!next)
2940                         goto out;
2941         }
2942
2943         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2944         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2945         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2946
2947         flags &= MAP_NONBLOCK;
2948         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2949         if (vma->vm_flags & VM_LOCKED)
2950                 flags |= MAP_LOCKED;
2951
2952         file = get_file(vma->vm_file);
2953         ret = do_mmap(vma->vm_file, start, size,
2954                         prot, flags, pgoff, &populate, NULL);
2955         fput(file);
2956 out:
2957         mmap_write_unlock(mm);
2958         if (populate)
2959                 mm_populate(ret, populate);
2960         if (!IS_ERR_VALUE(ret))
2961                 ret = 0;
2962         return ret;
2963 }
2964
2965 /*
2966  *  this is really a simplified "do_mmap".  it only handles
2967  *  anonymous maps.  eventually we may be able to do some
2968  *  brk-specific accounting here.
2969  */
2970 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
2971 {
2972         struct mm_struct *mm = current->mm;
2973         struct vm_area_struct *vma, *prev;
2974         struct rb_node **rb_link, *rb_parent;
2975         pgoff_t pgoff = addr >> PAGE_SHIFT;
2976         int error;
2977         unsigned long mapped_addr;
2978
2979         /* Until we need other flags, refuse anything except VM_EXEC. */
2980         if ((flags & (~VM_EXEC)) != 0)
2981                 return -EINVAL;
2982         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2983
2984         mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2985         if (IS_ERR_VALUE(mapped_addr))
2986                 return mapped_addr;
2987
2988         error = mlock_future_check(mm, mm->def_flags, len);
2989         if (error)
2990                 return error;
2991
2992         /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
2993         if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
2994                 return -ENOMEM;
2995
2996         /* Check against address space limits *after* clearing old maps... */
2997         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2998                 return -ENOMEM;
2999
3000         if (mm->map_count > sysctl_max_map_count)
3001                 return -ENOMEM;
3002
3003         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3004                 return -ENOMEM;
3005
3006         /* Can we just expand an old private anonymous mapping? */
3007         vma = vma_merge(mm, prev, addr, addr + len, flags,
3008                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
3009         if (vma)
3010                 goto out;
3011
3012         /*
3013          * create a vma struct for an anonymous mapping
3014          */
3015         vma = vm_area_alloc(mm);
3016         if (!vma) {
3017                 vm_unacct_memory(len >> PAGE_SHIFT);
3018                 return -ENOMEM;
3019         }
3020
3021         vma_set_anonymous(vma);
3022         vma->vm_start = addr;
3023         vma->vm_end = addr + len;
3024         vma->vm_pgoff = pgoff;
3025         vma->vm_flags = flags;
3026         vma->vm_page_prot = vm_get_page_prot(flags);
3027         vma_link(mm, vma, prev, rb_link, rb_parent);
3028 out:
3029         perf_event_mmap(vma);
3030         mm->total_vm += len >> PAGE_SHIFT;
3031         mm->data_vm += len >> PAGE_SHIFT;
3032         if (flags & VM_LOCKED)
3033                 mm->locked_vm += (len >> PAGE_SHIFT);
3034         vma->vm_flags |= VM_SOFTDIRTY;
3035         return 0;
3036 }
3037
3038 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3039 {
3040         struct mm_struct *mm = current->mm;
3041         unsigned long len;
3042         int ret;
3043         bool populate;
3044         LIST_HEAD(uf);
3045
3046         len = PAGE_ALIGN(request);
3047         if (len < request)
3048                 return -ENOMEM;
3049         if (!len)
3050                 return 0;
3051
3052         if (mmap_write_lock_killable(mm))
3053                 return -EINTR;
3054
3055         ret = do_brk_flags(addr, len, flags, &uf);
3056         populate = ((mm->def_flags & VM_LOCKED) != 0);
3057         mmap_write_unlock(mm);
3058         userfaultfd_unmap_complete(mm, &uf);
3059         if (populate && !ret)
3060                 mm_populate(addr, len);
3061         return ret;
3062 }
3063 EXPORT_SYMBOL(vm_brk_flags);
3064
3065 int vm_brk(unsigned long addr, unsigned long len)
3066 {
3067         return vm_brk_flags(addr, len, 0);
3068 }
3069 EXPORT_SYMBOL(vm_brk);
3070
3071 /* Release all mmaps. */
3072 void exit_mmap(struct mm_struct *mm)
3073 {
3074         struct mmu_gather tlb;
3075         struct vm_area_struct *vma;
3076         unsigned long nr_accounted = 0;
3077
3078         /* mm's last user has gone, and its about to be pulled down */
3079         mmu_notifier_release(mm);
3080
3081         if (unlikely(mm_is_oom_victim(mm))) {
3082                 /*
3083                  * Manually reap the mm to free as much memory as possible.
3084                  * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3085                  * this mm from further consideration.  Taking mm->mmap_lock for
3086                  * write after setting MMF_OOM_SKIP will guarantee that the oom
3087                  * reaper will not run on this mm again after mmap_lock is
3088                  * dropped.
3089                  *
3090                  * Nothing can be holding mm->mmap_lock here and the above call
3091                  * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3092                  * __oom_reap_task_mm() will not block.
3093                  */
3094                 (void)__oom_reap_task_mm(mm);
3095                 set_bit(MMF_OOM_SKIP, &mm->flags);
3096         }
3097
3098         mmap_write_lock(mm);
3099         arch_exit_mmap(mm);
3100
3101         vma = mm->mmap;
3102         if (!vma) {
3103                 /* Can happen if dup_mmap() received an OOM */
3104                 mmap_write_unlock(mm);
3105                 return;
3106         }
3107
3108         lru_add_drain();
3109         flush_cache_mm(mm);
3110         tlb_gather_mmu_fullmm(&tlb, mm);
3111         /* update_hiwater_rss(mm) here? but nobody should be looking */
3112         /* Use -1 here to ensure all VMAs in the mm are unmapped */
3113         unmap_vmas(&tlb, vma, 0, -1);
3114         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3115         tlb_finish_mmu(&tlb);
3116
3117         /* Walk the list again, actually closing and freeing it. */
3118         while (vma) {
3119                 if (vma->vm_flags & VM_ACCOUNT)
3120                         nr_accounted += vma_pages(vma);
3121                 vma = remove_vma(vma);
3122                 cond_resched();
3123         }
3124         mm->mmap = NULL;
3125         mmap_write_unlock(mm);
3126         vm_unacct_memory(nr_accounted);
3127 }
3128
3129 /* Insert vm structure into process list sorted by address
3130  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3131  * then i_mmap_rwsem is taken here.
3132  */
3133 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3134 {
3135         struct vm_area_struct *prev;
3136         struct rb_node **rb_link, *rb_parent;
3137
3138         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3139                            &prev, &rb_link, &rb_parent))
3140                 return -ENOMEM;
3141         if ((vma->vm_flags & VM_ACCOUNT) &&
3142              security_vm_enough_memory_mm(mm, vma_pages(vma)))
3143                 return -ENOMEM;
3144
3145         /*
3146          * The vm_pgoff of a purely anonymous vma should be irrelevant
3147          * until its first write fault, when page's anon_vma and index
3148          * are set.  But now set the vm_pgoff it will almost certainly
3149          * end up with (unless mremap moves it elsewhere before that
3150          * first wfault), so /proc/pid/maps tells a consistent story.
3151          *
3152          * By setting it to reflect the virtual start address of the
3153          * vma, merges and splits can happen in a seamless way, just
3154          * using the existing file pgoff checks and manipulations.
3155          * Similarly in do_mmap and in do_brk_flags.
3156          */
3157         if (vma_is_anonymous(vma)) {
3158                 BUG_ON(vma->anon_vma);
3159                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3160         }
3161
3162         vma_link(mm, vma, prev, rb_link, rb_parent);
3163         return 0;
3164 }
3165
3166 /*
3167  * Copy the vma structure to a new location in the same mm,
3168  * prior to moving page table entries, to effect an mremap move.
3169  */
3170 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3171         unsigned long addr, unsigned long len, pgoff_t pgoff,
3172         bool *need_rmap_locks)
3173 {
3174         struct vm_area_struct *vma = *vmap;
3175         unsigned long vma_start = vma->vm_start;
3176         struct mm_struct *mm = vma->vm_mm;
3177         struct vm_area_struct *new_vma, *prev;
3178         struct rb_node **rb_link, *rb_parent;
3179         bool faulted_in_anon_vma = true;
3180
3181         /*
3182          * If anonymous vma has not yet been faulted, update new pgoff
3183          * to match new location, to increase its chance of merging.
3184          */
3185         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3186                 pgoff = addr >> PAGE_SHIFT;
3187                 faulted_in_anon_vma = false;
3188         }
3189
3190         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3191                 return NULL;    /* should never get here */
3192         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3193                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3194                             vma->vm_userfaultfd_ctx, anon_vma_name(vma));
3195         if (new_vma) {
3196                 /*
3197                  * Source vma may have been merged into new_vma
3198                  */
3199                 if (unlikely(vma_start >= new_vma->vm_start &&
3200                              vma_start < new_vma->vm_end)) {
3201                         /*
3202                          * The only way we can get a vma_merge with
3203                          * self during an mremap is if the vma hasn't
3204                          * been faulted in yet and we were allowed to
3205                          * reset the dst vma->vm_pgoff to the
3206                          * destination address of the mremap to allow
3207                          * the merge to happen. mremap must change the
3208                          * vm_pgoff linearity between src and dst vmas
3209                          * (in turn preventing a vma_merge) to be
3210                          * safe. It is only safe to keep the vm_pgoff
3211                          * linear if there are no pages mapped yet.
3212                          */
3213                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3214                         *vmap = vma = new_vma;
3215                 }
3216                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3217         } else {
3218                 new_vma = vm_area_dup(vma);
3219                 if (!new_vma)
3220                         goto out;
3221                 new_vma->vm_start = addr;
3222                 new_vma->vm_end = addr + len;
3223                 new_vma->vm_pgoff = pgoff;
3224                 if (vma_dup_policy(vma, new_vma))
3225                         goto out_free_vma;
3226                 if (anon_vma_clone(new_vma, vma))
3227                         goto out_free_mempol;
3228                 if (new_vma->vm_file)
3229                         get_file(new_vma->vm_file);
3230                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3231                         new_vma->vm_ops->open(new_vma);
3232                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3233                 *need_rmap_locks = false;
3234         }
3235         return new_vma;
3236
3237 out_free_mempol:
3238         mpol_put(vma_policy(new_vma));
3239 out_free_vma:
3240         vm_area_free(new_vma);
3241 out:
3242         return NULL;
3243 }
3244
3245 /*
3246  * Return true if the calling process may expand its vm space by the passed
3247  * number of pages
3248  */
3249 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3250 {
3251         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3252                 return false;
3253
3254         if (is_data_mapping(flags) &&
3255             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3256                 /* Workaround for Valgrind */
3257                 if (rlimit(RLIMIT_DATA) == 0 &&
3258                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3259                         return true;
3260
3261                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3262                              current->comm, current->pid,
3263                              (mm->data_vm + npages) << PAGE_SHIFT,
3264                              rlimit(RLIMIT_DATA),
3265                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3266
3267                 if (!ignore_rlimit_data)
3268                         return false;
3269         }
3270
3271         return true;
3272 }
3273
3274 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3275 {
3276         WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3277
3278         if (is_exec_mapping(flags))
3279                 mm->exec_vm += npages;
3280         else if (is_stack_mapping(flags))
3281                 mm->stack_vm += npages;
3282         else if (is_data_mapping(flags))
3283                 mm->data_vm += npages;
3284 }
3285
3286 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3287
3288 /*
3289  * Having a close hook prevents vma merging regardless of flags.
3290  */
3291 static void special_mapping_close(struct vm_area_struct *vma)
3292 {
3293 }
3294
3295 static const char *special_mapping_name(struct vm_area_struct *vma)
3296 {
3297         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3298 }
3299
3300 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3301 {
3302         struct vm_special_mapping *sm = new_vma->vm_private_data;
3303
3304         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3305                 return -EFAULT;
3306
3307         if (sm->mremap)
3308                 return sm->mremap(sm, new_vma);
3309
3310         return 0;
3311 }
3312
3313 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3314 {
3315         /*
3316          * Forbid splitting special mappings - kernel has expectations over
3317          * the number of pages in mapping. Together with VM_DONTEXPAND
3318          * the size of vma should stay the same over the special mapping's
3319          * lifetime.
3320          */
3321         return -EINVAL;
3322 }
3323
3324 static const struct vm_operations_struct special_mapping_vmops = {
3325         .close = special_mapping_close,
3326         .fault = special_mapping_fault,
3327         .mremap = special_mapping_mremap,
3328         .name = special_mapping_name,
3329         /* vDSO code relies that VVAR can't be accessed remotely */
3330         .access = NULL,
3331         .may_split = special_mapping_split,
3332 };
3333
3334 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3335         .close = special_mapping_close,
3336         .fault = special_mapping_fault,
3337 };
3338
3339 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3340 {
3341         struct vm_area_struct *vma = vmf->vma;
3342         pgoff_t pgoff;
3343         struct page **pages;
3344
3345         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3346                 pages = vma->vm_private_data;
3347         } else {
3348                 struct vm_special_mapping *sm = vma->vm_private_data;
3349
3350                 if (sm->fault)
3351                         return sm->fault(sm, vmf->vma, vmf);
3352
3353                 pages = sm->pages;
3354         }
3355
3356         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3357                 pgoff--;
3358
3359         if (*pages) {
3360                 struct page *page = *pages;
3361                 get_page(page);
3362                 vmf->page = page;
3363                 return 0;
3364         }
3365
3366         return VM_FAULT_SIGBUS;
3367 }
3368
3369 static struct vm_area_struct *__install_special_mapping(
3370         struct mm_struct *mm,
3371         unsigned long addr, unsigned long len,
3372         unsigned long vm_flags, void *priv,
3373         const struct vm_operations_struct *ops)
3374 {
3375         int ret;
3376         struct vm_area_struct *vma;
3377
3378         vma = vm_area_alloc(mm);
3379         if (unlikely(vma == NULL))
3380                 return ERR_PTR(-ENOMEM);
3381
3382         vma->vm_start = addr;
3383         vma->vm_end = addr + len;
3384
3385         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3386         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
3387         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3388
3389         vma->vm_ops = ops;
3390         vma->vm_private_data = priv;
3391
3392         ret = insert_vm_struct(mm, vma);
3393         if (ret)
3394                 goto out;
3395
3396         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3397
3398         perf_event_mmap(vma);
3399
3400         return vma;
3401
3402 out:
3403         vm_area_free(vma);
3404         return ERR_PTR(ret);
3405 }
3406
3407 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3408         const struct vm_special_mapping *sm)
3409 {
3410         return vma->vm_private_data == sm &&
3411                 (vma->vm_ops == &special_mapping_vmops ||
3412                  vma->vm_ops == &legacy_special_mapping_vmops);
3413 }
3414
3415 /*
3416  * Called with mm->mmap_lock held for writing.
3417  * Insert a new vma covering the given region, with the given flags.
3418  * Its pages are supplied by the given array of struct page *.
3419  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3420  * The region past the last page supplied will always produce SIGBUS.
3421  * The array pointer and the pages it points to are assumed to stay alive
3422  * for as long as this mapping might exist.
3423  */
3424 struct vm_area_struct *_install_special_mapping(
3425         struct mm_struct *mm,
3426         unsigned long addr, unsigned long len,
3427         unsigned long vm_flags, const struct vm_special_mapping *spec)
3428 {
3429         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3430                                         &special_mapping_vmops);
3431 }
3432
3433 int install_special_mapping(struct mm_struct *mm,
3434                             unsigned long addr, unsigned long len,
3435                             unsigned long vm_flags, struct page **pages)
3436 {
3437         struct vm_area_struct *vma = __install_special_mapping(
3438                 mm, addr, len, vm_flags, (void *)pages,
3439                 &legacy_special_mapping_vmops);
3440
3441         return PTR_ERR_OR_ZERO(vma);
3442 }
3443
3444 static DEFINE_MUTEX(mm_all_locks_mutex);
3445
3446 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3447 {
3448         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3449                 /*
3450                  * The LSB of head.next can't change from under us
3451                  * because we hold the mm_all_locks_mutex.
3452                  */
3453                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3454                 /*
3455                  * We can safely modify head.next after taking the
3456                  * anon_vma->root->rwsem. If some other vma in this mm shares
3457                  * the same anon_vma we won't take it again.
3458                  *
3459                  * No need of atomic instructions here, head.next
3460                  * can't change from under us thanks to the
3461                  * anon_vma->root->rwsem.
3462                  */
3463                 if (__test_and_set_bit(0, (unsigned long *)
3464                                        &anon_vma->root->rb_root.rb_root.rb_node))
3465                         BUG();
3466         }
3467 }
3468
3469 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3470 {
3471         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3472                 /*
3473                  * AS_MM_ALL_LOCKS can't change from under us because
3474                  * we hold the mm_all_locks_mutex.
3475                  *
3476                  * Operations on ->flags have to be atomic because
3477                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3478                  * mm_all_locks_mutex, there may be other cpus
3479                  * changing other bitflags in parallel to us.
3480                  */
3481                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3482                         BUG();
3483                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3484         }
3485 }
3486
3487 /*
3488  * This operation locks against the VM for all pte/vma/mm related
3489  * operations that could ever happen on a certain mm. This includes
3490  * vmtruncate, try_to_unmap, and all page faults.
3491  *
3492  * The caller must take the mmap_lock in write mode before calling
3493  * mm_take_all_locks(). The caller isn't allowed to release the
3494  * mmap_lock until mm_drop_all_locks() returns.
3495  *
3496  * mmap_lock in write mode is required in order to block all operations
3497  * that could modify pagetables and free pages without need of
3498  * altering the vma layout. It's also needed in write mode to avoid new
3499  * anon_vmas to be associated with existing vmas.
3500  *
3501  * A single task can't take more than one mm_take_all_locks() in a row
3502  * or it would deadlock.
3503  *
3504  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3505  * mapping->flags avoid to take the same lock twice, if more than one
3506  * vma in this mm is backed by the same anon_vma or address_space.
3507  *
3508  * We take locks in following order, accordingly to comment at beginning
3509  * of mm/rmap.c:
3510  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3511  *     hugetlb mapping);
3512  *   - all i_mmap_rwsem locks;
3513  *   - all anon_vma->rwseml
3514  *
3515  * We can take all locks within these types randomly because the VM code
3516  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3517  * mm_all_locks_mutex.
3518  *
3519  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3520  * that may have to take thousand of locks.
3521  *
3522  * mm_take_all_locks() can fail if it's interrupted by signals.
3523  */
3524 int mm_take_all_locks(struct mm_struct *mm)
3525 {
3526         struct vm_area_struct *vma;
3527         struct anon_vma_chain *avc;
3528
3529         mmap_assert_write_locked(mm);
3530
3531         mutex_lock(&mm_all_locks_mutex);
3532
3533         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3534                 if (signal_pending(current))
3535                         goto out_unlock;
3536                 if (vma->vm_file && vma->vm_file->f_mapping &&
3537                                 is_vm_hugetlb_page(vma))
3538                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3539         }
3540
3541         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3542                 if (signal_pending(current))
3543                         goto out_unlock;
3544                 if (vma->vm_file && vma->vm_file->f_mapping &&
3545                                 !is_vm_hugetlb_page(vma))
3546                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3547         }
3548
3549         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3550                 if (signal_pending(current))
3551                         goto out_unlock;
3552                 if (vma->anon_vma)
3553                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3554                                 vm_lock_anon_vma(mm, avc->anon_vma);
3555         }
3556
3557         return 0;
3558
3559 out_unlock:
3560         mm_drop_all_locks(mm);
3561         return -EINTR;
3562 }
3563
3564 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3565 {
3566         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3567                 /*
3568                  * The LSB of head.next can't change to 0 from under
3569                  * us because we hold the mm_all_locks_mutex.
3570                  *
3571                  * We must however clear the bitflag before unlocking
3572                  * the vma so the users using the anon_vma->rb_root will
3573                  * never see our bitflag.
3574                  *
3575                  * No need of atomic instructions here, head.next
3576                  * can't change from under us until we release the
3577                  * anon_vma->root->rwsem.
3578                  */
3579                 if (!__test_and_clear_bit(0, (unsigned long *)
3580                                           &anon_vma->root->rb_root.rb_root.rb_node))
3581                         BUG();
3582                 anon_vma_unlock_write(anon_vma);
3583         }
3584 }
3585
3586 static void vm_unlock_mapping(struct address_space *mapping)
3587 {
3588         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3589                 /*
3590                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3591                  * because we hold the mm_all_locks_mutex.
3592                  */
3593                 i_mmap_unlock_write(mapping);
3594                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3595                                         &mapping->flags))
3596                         BUG();
3597         }
3598 }
3599
3600 /*
3601  * The mmap_lock cannot be released by the caller until
3602  * mm_drop_all_locks() returns.
3603  */
3604 void mm_drop_all_locks(struct mm_struct *mm)
3605 {
3606         struct vm_area_struct *vma;
3607         struct anon_vma_chain *avc;
3608
3609         mmap_assert_write_locked(mm);
3610         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3611
3612         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3613                 if (vma->anon_vma)
3614                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3615                                 vm_unlock_anon_vma(avc->anon_vma);
3616                 if (vma->vm_file && vma->vm_file->f_mapping)
3617                         vm_unlock_mapping(vma->vm_file->f_mapping);
3618         }
3619
3620         mutex_unlock(&mm_all_locks_mutex);
3621 }
3622
3623 /*
3624  * initialise the percpu counter for VM
3625  */
3626 void __init mmap_init(void)
3627 {
3628         int ret;
3629
3630         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3631         VM_BUG_ON(ret);
3632 }
3633
3634 /*
3635  * Initialise sysctl_user_reserve_kbytes.
3636  *
3637  * This is intended to prevent a user from starting a single memory hogging
3638  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3639  * mode.
3640  *
3641  * The default value is min(3% of free memory, 128MB)
3642  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3643  */
3644 static int init_user_reserve(void)
3645 {
3646         unsigned long free_kbytes;
3647
3648         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3649
3650         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3651         return 0;
3652 }
3653 subsys_initcall(init_user_reserve);
3654
3655 /*
3656  * Initialise sysctl_admin_reserve_kbytes.
3657  *
3658  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3659  * to log in and kill a memory hogging process.
3660  *
3661  * Systems with more than 256MB will reserve 8MB, enough to recover
3662  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3663  * only reserve 3% of free pages by default.
3664  */
3665 static int init_admin_reserve(void)
3666 {
3667         unsigned long free_kbytes;
3668
3669         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3670
3671         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3672         return 0;
3673 }
3674 subsys_initcall(init_admin_reserve);
3675
3676 /*
3677  * Reinititalise user and admin reserves if memory is added or removed.
3678  *
3679  * The default user reserve max is 128MB, and the default max for the
3680  * admin reserve is 8MB. These are usually, but not always, enough to
3681  * enable recovery from a memory hogging process using login/sshd, a shell,
3682  * and tools like top. It may make sense to increase or even disable the
3683  * reserve depending on the existence of swap or variations in the recovery
3684  * tools. So, the admin may have changed them.
3685  *
3686  * If memory is added and the reserves have been eliminated or increased above
3687  * the default max, then we'll trust the admin.
3688  *
3689  * If memory is removed and there isn't enough free memory, then we
3690  * need to reset the reserves.
3691  *
3692  * Otherwise keep the reserve set by the admin.
3693  */
3694 static int reserve_mem_notifier(struct notifier_block *nb,
3695                              unsigned long action, void *data)
3696 {
3697         unsigned long tmp, free_kbytes;
3698
3699         switch (action) {
3700         case MEM_ONLINE:
3701                 /* Default max is 128MB. Leave alone if modified by operator. */
3702                 tmp = sysctl_user_reserve_kbytes;
3703                 if (0 < tmp && tmp < (1UL << 17))
3704                         init_user_reserve();
3705
3706                 /* Default max is 8MB.  Leave alone if modified by operator. */
3707                 tmp = sysctl_admin_reserve_kbytes;
3708                 if (0 < tmp && tmp < (1UL << 13))
3709                         init_admin_reserve();
3710
3711                 break;
3712         case MEM_OFFLINE:
3713                 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3714
3715                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3716                         init_user_reserve();
3717                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3718                                 sysctl_user_reserve_kbytes);
3719                 }
3720
3721                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3722                         init_admin_reserve();
3723                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3724                                 sysctl_admin_reserve_kbytes);
3725                 }
3726                 break;
3727         default:
3728                 break;
3729         }
3730         return NOTIFY_OK;
3731 }
3732
3733 static struct notifier_block reserve_mem_nb = {
3734         .notifier_call = reserve_mem_notifier,
3735 };
3736
3737 static int __meminit init_reserve_notifier(void)
3738 {
3739         if (register_hotmemory_notifier(&reserve_mem_nb))
3740                 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3741
3742         return 0;
3743 }
3744 subsys_initcall(init_reserve_notifier);