Merge tag 'drm-misc-next-fixes-2020-04-04' of git://anongit.freedesktop.org/drm/drm...
[linux-2.6-microblaze.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/vmacache.h>
17 #include <linux/shm.h>
18 #include <linux/mman.h>
19 #include <linux/pagemap.h>
20 #include <linux/swap.h>
21 #include <linux/syscalls.h>
22 #include <linux/capability.h>
23 #include <linux/init.h>
24 #include <linux/file.h>
25 #include <linux/fs.h>
26 #include <linux/personality.h>
27 #include <linux/security.h>
28 #include <linux/hugetlb.h>
29 #include <linux/shmem_fs.h>
30 #include <linux/profile.h>
31 #include <linux/export.h>
32 #include <linux/mount.h>
33 #include <linux/mempolicy.h>
34 #include <linux/rmap.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/mmdebug.h>
37 #include <linux/perf_event.h>
38 #include <linux/audit.h>
39 #include <linux/khugepaged.h>
40 #include <linux/uprobes.h>
41 #include <linux/rbtree_augmented.h>
42 #include <linux/notifier.h>
43 #include <linux/memory.h>
44 #include <linux/printk.h>
45 #include <linux/userfaultfd_k.h>
46 #include <linux/moduleparam.h>
47 #include <linux/pkeys.h>
48 #include <linux/oom.h>
49 #include <linux/sched/mm.h>
50
51 #include <linux/uaccess.h>
52 #include <asm/cacheflush.h>
53 #include <asm/tlb.h>
54 #include <asm/mmu_context.h>
55
56 #include "internal.h"
57
58 #ifndef arch_mmap_check
59 #define arch_mmap_check(addr, len, flags)       (0)
60 #endif
61
62 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
63 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
64 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
65 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
66 #endif
67 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
68 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
69 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
70 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
71 #endif
72
73 static bool ignore_rlimit_data;
74 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
75
76 static void unmap_region(struct mm_struct *mm,
77                 struct vm_area_struct *vma, struct vm_area_struct *prev,
78                 unsigned long start, unsigned long end);
79
80 /* description of effects of mapping type and prot in current implementation.
81  * this is due to the limited x86 page protection hardware.  The expected
82  * behavior is in parens:
83  *
84  * map_type     prot
85  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
86  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
87  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
88  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
89  *
90  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
91  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
92  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
93  */
94 pgprot_t protection_map[16] __ro_after_init = {
95         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
96         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
97 };
98
99 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
100 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
101 {
102         return prot;
103 }
104 #endif
105
106 pgprot_t vm_get_page_prot(unsigned long vm_flags)
107 {
108         pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
109                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
110                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
111
112         return arch_filter_pgprot(ret);
113 }
114 EXPORT_SYMBOL(vm_get_page_prot);
115
116 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
117 {
118         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
119 }
120
121 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
122 void vma_set_page_prot(struct vm_area_struct *vma)
123 {
124         unsigned long vm_flags = vma->vm_flags;
125         pgprot_t vm_page_prot;
126
127         vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
128         if (vma_wants_writenotify(vma, vm_page_prot)) {
129                 vm_flags &= ~VM_SHARED;
130                 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
131         }
132         /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
133         WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
134 }
135
136 /*
137  * Requires inode->i_mapping->i_mmap_rwsem
138  */
139 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
140                 struct file *file, struct address_space *mapping)
141 {
142         if (vma->vm_flags & VM_DENYWRITE)
143                 atomic_inc(&file_inode(file)->i_writecount);
144         if (vma->vm_flags & VM_SHARED)
145                 mapping_unmap_writable(mapping);
146
147         flush_dcache_mmap_lock(mapping);
148         vma_interval_tree_remove(vma, &mapping->i_mmap);
149         flush_dcache_mmap_unlock(mapping);
150 }
151
152 /*
153  * Unlink a file-based vm structure from its interval tree, to hide
154  * vma from rmap and vmtruncate before freeing its page tables.
155  */
156 void unlink_file_vma(struct vm_area_struct *vma)
157 {
158         struct file *file = vma->vm_file;
159
160         if (file) {
161                 struct address_space *mapping = file->f_mapping;
162                 i_mmap_lock_write(mapping);
163                 __remove_shared_vm_struct(vma, file, mapping);
164                 i_mmap_unlock_write(mapping);
165         }
166 }
167
168 /*
169  * Close a vm structure and free it, returning the next.
170  */
171 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
172 {
173         struct vm_area_struct *next = vma->vm_next;
174
175         might_sleep();
176         if (vma->vm_ops && vma->vm_ops->close)
177                 vma->vm_ops->close(vma);
178         if (vma->vm_file)
179                 fput(vma->vm_file);
180         mpol_put(vma_policy(vma));
181         vm_area_free(vma);
182         return next;
183 }
184
185 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
186                 struct list_head *uf);
187 SYSCALL_DEFINE1(brk, unsigned long, brk)
188 {
189         unsigned long retval;
190         unsigned long newbrk, oldbrk, origbrk;
191         struct mm_struct *mm = current->mm;
192         struct vm_area_struct *next;
193         unsigned long min_brk;
194         bool populate;
195         bool downgraded = false;
196         LIST_HEAD(uf);
197
198         if (down_write_killable(&mm->mmap_sem))
199                 return -EINTR;
200
201         origbrk = mm->brk;
202
203 #ifdef CONFIG_COMPAT_BRK
204         /*
205          * CONFIG_COMPAT_BRK can still be overridden by setting
206          * randomize_va_space to 2, which will still cause mm->start_brk
207          * to be arbitrarily shifted
208          */
209         if (current->brk_randomized)
210                 min_brk = mm->start_brk;
211         else
212                 min_brk = mm->end_data;
213 #else
214         min_brk = mm->start_brk;
215 #endif
216         if (brk < min_brk)
217                 goto out;
218
219         /*
220          * Check against rlimit here. If this check is done later after the test
221          * of oldbrk with newbrk then it can escape the test and let the data
222          * segment grow beyond its set limit the in case where the limit is
223          * not page aligned -Ram Gupta
224          */
225         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
226                               mm->end_data, mm->start_data))
227                 goto out;
228
229         newbrk = PAGE_ALIGN(brk);
230         oldbrk = PAGE_ALIGN(mm->brk);
231         if (oldbrk == newbrk) {
232                 mm->brk = brk;
233                 goto success;
234         }
235
236         /*
237          * Always allow shrinking brk.
238          * __do_munmap() may downgrade mmap_sem to read.
239          */
240         if (brk <= mm->brk) {
241                 int ret;
242
243                 /*
244                  * mm->brk must to be protected by write mmap_sem so update it
245                  * before downgrading mmap_sem. When __do_munmap() fails,
246                  * mm->brk will be restored from origbrk.
247                  */
248                 mm->brk = brk;
249                 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
250                 if (ret < 0) {
251                         mm->brk = origbrk;
252                         goto out;
253                 } else if (ret == 1) {
254                         downgraded = true;
255                 }
256                 goto success;
257         }
258
259         /* Check against existing mmap mappings. */
260         next = find_vma(mm, oldbrk);
261         if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
262                 goto out;
263
264         /* Ok, looks good - let it rip. */
265         if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
266                 goto out;
267         mm->brk = brk;
268
269 success:
270         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
271         if (downgraded)
272                 up_read(&mm->mmap_sem);
273         else
274                 up_write(&mm->mmap_sem);
275         userfaultfd_unmap_complete(mm, &uf);
276         if (populate)
277                 mm_populate(oldbrk, newbrk - oldbrk);
278         return brk;
279
280 out:
281         retval = origbrk;
282         up_write(&mm->mmap_sem);
283         return retval;
284 }
285
286 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
287 {
288         unsigned long gap, prev_end;
289
290         /*
291          * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
292          * allow two stack_guard_gaps between them here, and when choosing
293          * an unmapped area; whereas when expanding we only require one.
294          * That's a little inconsistent, but keeps the code here simpler.
295          */
296         gap = vm_start_gap(vma);
297         if (vma->vm_prev) {
298                 prev_end = vm_end_gap(vma->vm_prev);
299                 if (gap > prev_end)
300                         gap -= prev_end;
301                 else
302                         gap = 0;
303         }
304         return gap;
305 }
306
307 #ifdef CONFIG_DEBUG_VM_RB
308 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
309 {
310         unsigned long max = vma_compute_gap(vma), subtree_gap;
311         if (vma->vm_rb.rb_left) {
312                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
313                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
314                 if (subtree_gap > max)
315                         max = subtree_gap;
316         }
317         if (vma->vm_rb.rb_right) {
318                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
319                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
320                 if (subtree_gap > max)
321                         max = subtree_gap;
322         }
323         return max;
324 }
325
326 static int browse_rb(struct mm_struct *mm)
327 {
328         struct rb_root *root = &mm->mm_rb;
329         int i = 0, j, bug = 0;
330         struct rb_node *nd, *pn = NULL;
331         unsigned long prev = 0, pend = 0;
332
333         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
334                 struct vm_area_struct *vma;
335                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
336                 if (vma->vm_start < prev) {
337                         pr_emerg("vm_start %lx < prev %lx\n",
338                                   vma->vm_start, prev);
339                         bug = 1;
340                 }
341                 if (vma->vm_start < pend) {
342                         pr_emerg("vm_start %lx < pend %lx\n",
343                                   vma->vm_start, pend);
344                         bug = 1;
345                 }
346                 if (vma->vm_start > vma->vm_end) {
347                         pr_emerg("vm_start %lx > vm_end %lx\n",
348                                   vma->vm_start, vma->vm_end);
349                         bug = 1;
350                 }
351                 spin_lock(&mm->page_table_lock);
352                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
353                         pr_emerg("free gap %lx, correct %lx\n",
354                                vma->rb_subtree_gap,
355                                vma_compute_subtree_gap(vma));
356                         bug = 1;
357                 }
358                 spin_unlock(&mm->page_table_lock);
359                 i++;
360                 pn = nd;
361                 prev = vma->vm_start;
362                 pend = vma->vm_end;
363         }
364         j = 0;
365         for (nd = pn; nd; nd = rb_prev(nd))
366                 j++;
367         if (i != j) {
368                 pr_emerg("backwards %d, forwards %d\n", j, i);
369                 bug = 1;
370         }
371         return bug ? -1 : i;
372 }
373
374 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
375 {
376         struct rb_node *nd;
377
378         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
379                 struct vm_area_struct *vma;
380                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
381                 VM_BUG_ON_VMA(vma != ignore &&
382                         vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
383                         vma);
384         }
385 }
386
387 static void validate_mm(struct mm_struct *mm)
388 {
389         int bug = 0;
390         int i = 0;
391         unsigned long highest_address = 0;
392         struct vm_area_struct *vma = mm->mmap;
393
394         while (vma) {
395                 struct anon_vma *anon_vma = vma->anon_vma;
396                 struct anon_vma_chain *avc;
397
398                 if (anon_vma) {
399                         anon_vma_lock_read(anon_vma);
400                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
401                                 anon_vma_interval_tree_verify(avc);
402                         anon_vma_unlock_read(anon_vma);
403                 }
404
405                 highest_address = vm_end_gap(vma);
406                 vma = vma->vm_next;
407                 i++;
408         }
409         if (i != mm->map_count) {
410                 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
411                 bug = 1;
412         }
413         if (highest_address != mm->highest_vm_end) {
414                 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
415                           mm->highest_vm_end, highest_address);
416                 bug = 1;
417         }
418         i = browse_rb(mm);
419         if (i != mm->map_count) {
420                 if (i != -1)
421                         pr_emerg("map_count %d rb %d\n", mm->map_count, i);
422                 bug = 1;
423         }
424         VM_BUG_ON_MM(bug, mm);
425 }
426 #else
427 #define validate_mm_rb(root, ignore) do { } while (0)
428 #define validate_mm(mm) do { } while (0)
429 #endif
430
431 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
432                          struct vm_area_struct, vm_rb,
433                          unsigned long, rb_subtree_gap, vma_compute_gap)
434
435 /*
436  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
437  * vma->vm_prev->vm_end values changed, without modifying the vma's position
438  * in the rbtree.
439  */
440 static void vma_gap_update(struct vm_area_struct *vma)
441 {
442         /*
443          * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
444          * a callback function that does exactly what we want.
445          */
446         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
447 }
448
449 static inline void vma_rb_insert(struct vm_area_struct *vma,
450                                  struct rb_root *root)
451 {
452         /* All rb_subtree_gap values must be consistent prior to insertion */
453         validate_mm_rb(root, NULL);
454
455         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
456 }
457
458 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
459 {
460         /*
461          * Note rb_erase_augmented is a fairly large inline function,
462          * so make sure we instantiate it only once with our desired
463          * augmented rbtree callbacks.
464          */
465         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
466 }
467
468 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
469                                                 struct rb_root *root,
470                                                 struct vm_area_struct *ignore)
471 {
472         /*
473          * All rb_subtree_gap values must be consistent prior to erase,
474          * with the possible exception of the "next" vma being erased if
475          * next->vm_start was reduced.
476          */
477         validate_mm_rb(root, ignore);
478
479         __vma_rb_erase(vma, root);
480 }
481
482 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
483                                          struct rb_root *root)
484 {
485         /*
486          * All rb_subtree_gap values must be consistent prior to erase,
487          * with the possible exception of the vma being erased.
488          */
489         validate_mm_rb(root, vma);
490
491         __vma_rb_erase(vma, root);
492 }
493
494 /*
495  * vma has some anon_vma assigned, and is already inserted on that
496  * anon_vma's interval trees.
497  *
498  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
499  * vma must be removed from the anon_vma's interval trees using
500  * anon_vma_interval_tree_pre_update_vma().
501  *
502  * After the update, the vma will be reinserted using
503  * anon_vma_interval_tree_post_update_vma().
504  *
505  * The entire update must be protected by exclusive mmap_sem and by
506  * the root anon_vma's mutex.
507  */
508 static inline void
509 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
510 {
511         struct anon_vma_chain *avc;
512
513         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
514                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
515 }
516
517 static inline void
518 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
519 {
520         struct anon_vma_chain *avc;
521
522         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
523                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
524 }
525
526 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
527                 unsigned long end, struct vm_area_struct **pprev,
528                 struct rb_node ***rb_link, struct rb_node **rb_parent)
529 {
530         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
531
532         __rb_link = &mm->mm_rb.rb_node;
533         rb_prev = __rb_parent = NULL;
534
535         while (*__rb_link) {
536                 struct vm_area_struct *vma_tmp;
537
538                 __rb_parent = *__rb_link;
539                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
540
541                 if (vma_tmp->vm_end > addr) {
542                         /* Fail if an existing vma overlaps the area */
543                         if (vma_tmp->vm_start < end)
544                                 return -ENOMEM;
545                         __rb_link = &__rb_parent->rb_left;
546                 } else {
547                         rb_prev = __rb_parent;
548                         __rb_link = &__rb_parent->rb_right;
549                 }
550         }
551
552         *pprev = NULL;
553         if (rb_prev)
554                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
555         *rb_link = __rb_link;
556         *rb_parent = __rb_parent;
557         return 0;
558 }
559
560 static unsigned long count_vma_pages_range(struct mm_struct *mm,
561                 unsigned long addr, unsigned long end)
562 {
563         unsigned long nr_pages = 0;
564         struct vm_area_struct *vma;
565
566         /* Find first overlaping mapping */
567         vma = find_vma_intersection(mm, addr, end);
568         if (!vma)
569                 return 0;
570
571         nr_pages = (min(end, vma->vm_end) -
572                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
573
574         /* Iterate over the rest of the overlaps */
575         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
576                 unsigned long overlap_len;
577
578                 if (vma->vm_start > end)
579                         break;
580
581                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
582                 nr_pages += overlap_len >> PAGE_SHIFT;
583         }
584
585         return nr_pages;
586 }
587
588 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
589                 struct rb_node **rb_link, struct rb_node *rb_parent)
590 {
591         /* Update tracking information for the gap following the new vma. */
592         if (vma->vm_next)
593                 vma_gap_update(vma->vm_next);
594         else
595                 mm->highest_vm_end = vm_end_gap(vma);
596
597         /*
598          * vma->vm_prev wasn't known when we followed the rbtree to find the
599          * correct insertion point for that vma. As a result, we could not
600          * update the vma vm_rb parents rb_subtree_gap values on the way down.
601          * So, we first insert the vma with a zero rb_subtree_gap value
602          * (to be consistent with what we did on the way down), and then
603          * immediately update the gap to the correct value. Finally we
604          * rebalance the rbtree after all augmented values have been set.
605          */
606         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
607         vma->rb_subtree_gap = 0;
608         vma_gap_update(vma);
609         vma_rb_insert(vma, &mm->mm_rb);
610 }
611
612 static void __vma_link_file(struct vm_area_struct *vma)
613 {
614         struct file *file;
615
616         file = vma->vm_file;
617         if (file) {
618                 struct address_space *mapping = file->f_mapping;
619
620                 if (vma->vm_flags & VM_DENYWRITE)
621                         atomic_dec(&file_inode(file)->i_writecount);
622                 if (vma->vm_flags & VM_SHARED)
623                         atomic_inc(&mapping->i_mmap_writable);
624
625                 flush_dcache_mmap_lock(mapping);
626                 vma_interval_tree_insert(vma, &mapping->i_mmap);
627                 flush_dcache_mmap_unlock(mapping);
628         }
629 }
630
631 static void
632 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
633         struct vm_area_struct *prev, struct rb_node **rb_link,
634         struct rb_node *rb_parent)
635 {
636         __vma_link_list(mm, vma, prev);
637         __vma_link_rb(mm, vma, rb_link, rb_parent);
638 }
639
640 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
641                         struct vm_area_struct *prev, struct rb_node **rb_link,
642                         struct rb_node *rb_parent)
643 {
644         struct address_space *mapping = NULL;
645
646         if (vma->vm_file) {
647                 mapping = vma->vm_file->f_mapping;
648                 i_mmap_lock_write(mapping);
649         }
650
651         __vma_link(mm, vma, prev, rb_link, rb_parent);
652         __vma_link_file(vma);
653
654         if (mapping)
655                 i_mmap_unlock_write(mapping);
656
657         mm->map_count++;
658         validate_mm(mm);
659 }
660
661 /*
662  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
663  * mm's list and rbtree.  It has already been inserted into the interval tree.
664  */
665 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
666 {
667         struct vm_area_struct *prev;
668         struct rb_node **rb_link, *rb_parent;
669
670         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
671                            &prev, &rb_link, &rb_parent))
672                 BUG();
673         __vma_link(mm, vma, prev, rb_link, rb_parent);
674         mm->map_count++;
675 }
676
677 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
678                                                 struct vm_area_struct *vma,
679                                                 struct vm_area_struct *ignore)
680 {
681         vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
682         __vma_unlink_list(mm, vma);
683         /* Kill the cache */
684         vmacache_invalidate(mm);
685 }
686
687 /*
688  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
689  * is already present in an i_mmap tree without adjusting the tree.
690  * The following helper function should be used when such adjustments
691  * are necessary.  The "insert" vma (if any) is to be inserted
692  * before we drop the necessary locks.
693  */
694 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
695         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
696         struct vm_area_struct *expand)
697 {
698         struct mm_struct *mm = vma->vm_mm;
699         struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
700         struct address_space *mapping = NULL;
701         struct rb_root_cached *root = NULL;
702         struct anon_vma *anon_vma = NULL;
703         struct file *file = vma->vm_file;
704         bool start_changed = false, end_changed = false;
705         long adjust_next = 0;
706         int remove_next = 0;
707
708         if (next && !insert) {
709                 struct vm_area_struct *exporter = NULL, *importer = NULL;
710
711                 if (end >= next->vm_end) {
712                         /*
713                          * vma expands, overlapping all the next, and
714                          * perhaps the one after too (mprotect case 6).
715                          * The only other cases that gets here are
716                          * case 1, case 7 and case 8.
717                          */
718                         if (next == expand) {
719                                 /*
720                                  * The only case where we don't expand "vma"
721                                  * and we expand "next" instead is case 8.
722                                  */
723                                 VM_WARN_ON(end != next->vm_end);
724                                 /*
725                                  * remove_next == 3 means we're
726                                  * removing "vma" and that to do so we
727                                  * swapped "vma" and "next".
728                                  */
729                                 remove_next = 3;
730                                 VM_WARN_ON(file != next->vm_file);
731                                 swap(vma, next);
732                         } else {
733                                 VM_WARN_ON(expand != vma);
734                                 /*
735                                  * case 1, 6, 7, remove_next == 2 is case 6,
736                                  * remove_next == 1 is case 1 or 7.
737                                  */
738                                 remove_next = 1 + (end > next->vm_end);
739                                 VM_WARN_ON(remove_next == 2 &&
740                                            end != next->vm_next->vm_end);
741                                 /* trim end to next, for case 6 first pass */
742                                 end = next->vm_end;
743                         }
744
745                         exporter = next;
746                         importer = vma;
747
748                         /*
749                          * If next doesn't have anon_vma, import from vma after
750                          * next, if the vma overlaps with it.
751                          */
752                         if (remove_next == 2 && !next->anon_vma)
753                                 exporter = next->vm_next;
754
755                 } else if (end > next->vm_start) {
756                         /*
757                          * vma expands, overlapping part of the next:
758                          * mprotect case 5 shifting the boundary up.
759                          */
760                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
761                         exporter = next;
762                         importer = vma;
763                         VM_WARN_ON(expand != importer);
764                 } else if (end < vma->vm_end) {
765                         /*
766                          * vma shrinks, and !insert tells it's not
767                          * split_vma inserting another: so it must be
768                          * mprotect case 4 shifting the boundary down.
769                          */
770                         adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
771                         exporter = vma;
772                         importer = next;
773                         VM_WARN_ON(expand != importer);
774                 }
775
776                 /*
777                  * Easily overlooked: when mprotect shifts the boundary,
778                  * make sure the expanding vma has anon_vma set if the
779                  * shrinking vma had, to cover any anon pages imported.
780                  */
781                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
782                         int error;
783
784                         importer->anon_vma = exporter->anon_vma;
785                         error = anon_vma_clone(importer, exporter);
786                         if (error)
787                                 return error;
788                 }
789         }
790 again:
791         vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
792
793         if (file) {
794                 mapping = file->f_mapping;
795                 root = &mapping->i_mmap;
796                 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
797
798                 if (adjust_next)
799                         uprobe_munmap(next, next->vm_start, next->vm_end);
800
801                 i_mmap_lock_write(mapping);
802                 if (insert) {
803                         /*
804                          * Put into interval tree now, so instantiated pages
805                          * are visible to arm/parisc __flush_dcache_page
806                          * throughout; but we cannot insert into address
807                          * space until vma start or end is updated.
808                          */
809                         __vma_link_file(insert);
810                 }
811         }
812
813         anon_vma = vma->anon_vma;
814         if (!anon_vma && adjust_next)
815                 anon_vma = next->anon_vma;
816         if (anon_vma) {
817                 VM_WARN_ON(adjust_next && next->anon_vma &&
818                            anon_vma != next->anon_vma);
819                 anon_vma_lock_write(anon_vma);
820                 anon_vma_interval_tree_pre_update_vma(vma);
821                 if (adjust_next)
822                         anon_vma_interval_tree_pre_update_vma(next);
823         }
824
825         if (root) {
826                 flush_dcache_mmap_lock(mapping);
827                 vma_interval_tree_remove(vma, root);
828                 if (adjust_next)
829                         vma_interval_tree_remove(next, root);
830         }
831
832         if (start != vma->vm_start) {
833                 vma->vm_start = start;
834                 start_changed = true;
835         }
836         if (end != vma->vm_end) {
837                 vma->vm_end = end;
838                 end_changed = true;
839         }
840         vma->vm_pgoff = pgoff;
841         if (adjust_next) {
842                 next->vm_start += adjust_next << PAGE_SHIFT;
843                 next->vm_pgoff += adjust_next;
844         }
845
846         if (root) {
847                 if (adjust_next)
848                         vma_interval_tree_insert(next, root);
849                 vma_interval_tree_insert(vma, root);
850                 flush_dcache_mmap_unlock(mapping);
851         }
852
853         if (remove_next) {
854                 /*
855                  * vma_merge has merged next into vma, and needs
856                  * us to remove next before dropping the locks.
857                  */
858                 if (remove_next != 3)
859                         __vma_unlink_common(mm, next, next);
860                 else
861                         /*
862                          * vma is not before next if they've been
863                          * swapped.
864                          *
865                          * pre-swap() next->vm_start was reduced so
866                          * tell validate_mm_rb to ignore pre-swap()
867                          * "next" (which is stored in post-swap()
868                          * "vma").
869                          */
870                         __vma_unlink_common(mm, next, vma);
871                 if (file)
872                         __remove_shared_vm_struct(next, file, mapping);
873         } else if (insert) {
874                 /*
875                  * split_vma has split insert from vma, and needs
876                  * us to insert it before dropping the locks
877                  * (it may either follow vma or precede it).
878                  */
879                 __insert_vm_struct(mm, insert);
880         } else {
881                 if (start_changed)
882                         vma_gap_update(vma);
883                 if (end_changed) {
884                         if (!next)
885                                 mm->highest_vm_end = vm_end_gap(vma);
886                         else if (!adjust_next)
887                                 vma_gap_update(next);
888                 }
889         }
890
891         if (anon_vma) {
892                 anon_vma_interval_tree_post_update_vma(vma);
893                 if (adjust_next)
894                         anon_vma_interval_tree_post_update_vma(next);
895                 anon_vma_unlock_write(anon_vma);
896         }
897         if (mapping)
898                 i_mmap_unlock_write(mapping);
899
900         if (root) {
901                 uprobe_mmap(vma);
902
903                 if (adjust_next)
904                         uprobe_mmap(next);
905         }
906
907         if (remove_next) {
908                 if (file) {
909                         uprobe_munmap(next, next->vm_start, next->vm_end);
910                         fput(file);
911                 }
912                 if (next->anon_vma)
913                         anon_vma_merge(vma, next);
914                 mm->map_count--;
915                 mpol_put(vma_policy(next));
916                 vm_area_free(next);
917                 /*
918                  * In mprotect's case 6 (see comments on vma_merge),
919                  * we must remove another next too. It would clutter
920                  * up the code too much to do both in one go.
921                  */
922                 if (remove_next != 3) {
923                         /*
924                          * If "next" was removed and vma->vm_end was
925                          * expanded (up) over it, in turn
926                          * "next->vm_prev->vm_end" changed and the
927                          * "vma->vm_next" gap must be updated.
928                          */
929                         next = vma->vm_next;
930                 } else {
931                         /*
932                          * For the scope of the comment "next" and
933                          * "vma" considered pre-swap(): if "vma" was
934                          * removed, next->vm_start was expanded (down)
935                          * over it and the "next" gap must be updated.
936                          * Because of the swap() the post-swap() "vma"
937                          * actually points to pre-swap() "next"
938                          * (post-swap() "next" as opposed is now a
939                          * dangling pointer).
940                          */
941                         next = vma;
942                 }
943                 if (remove_next == 2) {
944                         remove_next = 1;
945                         end = next->vm_end;
946                         goto again;
947                 }
948                 else if (next)
949                         vma_gap_update(next);
950                 else {
951                         /*
952                          * If remove_next == 2 we obviously can't
953                          * reach this path.
954                          *
955                          * If remove_next == 3 we can't reach this
956                          * path because pre-swap() next is always not
957                          * NULL. pre-swap() "next" is not being
958                          * removed and its next->vm_end is not altered
959                          * (and furthermore "end" already matches
960                          * next->vm_end in remove_next == 3).
961                          *
962                          * We reach this only in the remove_next == 1
963                          * case if the "next" vma that was removed was
964                          * the highest vma of the mm. However in such
965                          * case next->vm_end == "end" and the extended
966                          * "vma" has vma->vm_end == next->vm_end so
967                          * mm->highest_vm_end doesn't need any update
968                          * in remove_next == 1 case.
969                          */
970                         VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
971                 }
972         }
973         if (insert && file)
974                 uprobe_mmap(insert);
975
976         validate_mm(mm);
977
978         return 0;
979 }
980
981 /*
982  * If the vma has a ->close operation then the driver probably needs to release
983  * per-vma resources, so we don't attempt to merge those.
984  */
985 static inline int is_mergeable_vma(struct vm_area_struct *vma,
986                                 struct file *file, unsigned long vm_flags,
987                                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
988 {
989         /*
990          * VM_SOFTDIRTY should not prevent from VMA merging, if we
991          * match the flags but dirty bit -- the caller should mark
992          * merged VMA as dirty. If dirty bit won't be excluded from
993          * comparison, we increase pressure on the memory system forcing
994          * the kernel to generate new VMAs when old one could be
995          * extended instead.
996          */
997         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
998                 return 0;
999         if (vma->vm_file != file)
1000                 return 0;
1001         if (vma->vm_ops && vma->vm_ops->close)
1002                 return 0;
1003         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1004                 return 0;
1005         return 1;
1006 }
1007
1008 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1009                                         struct anon_vma *anon_vma2,
1010                                         struct vm_area_struct *vma)
1011 {
1012         /*
1013          * The list_is_singular() test is to avoid merging VMA cloned from
1014          * parents. This can improve scalability caused by anon_vma lock.
1015          */
1016         if ((!anon_vma1 || !anon_vma2) && (!vma ||
1017                 list_is_singular(&vma->anon_vma_chain)))
1018                 return 1;
1019         return anon_vma1 == anon_vma2;
1020 }
1021
1022 /*
1023  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1024  * in front of (at a lower virtual address and file offset than) the vma.
1025  *
1026  * We cannot merge two vmas if they have differently assigned (non-NULL)
1027  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1028  *
1029  * We don't check here for the merged mmap wrapping around the end of pagecache
1030  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1031  * wrap, nor mmaps which cover the final page at index -1UL.
1032  */
1033 static int
1034 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1035                      struct anon_vma *anon_vma, struct file *file,
1036                      pgoff_t vm_pgoff,
1037                      struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1038 {
1039         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1040             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1041                 if (vma->vm_pgoff == vm_pgoff)
1042                         return 1;
1043         }
1044         return 0;
1045 }
1046
1047 /*
1048  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1049  * beyond (at a higher virtual address and file offset than) the vma.
1050  *
1051  * We cannot merge two vmas if they have differently assigned (non-NULL)
1052  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1053  */
1054 static int
1055 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1056                     struct anon_vma *anon_vma, struct file *file,
1057                     pgoff_t vm_pgoff,
1058                     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1059 {
1060         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1061             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1062                 pgoff_t vm_pglen;
1063                 vm_pglen = vma_pages(vma);
1064                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1065                         return 1;
1066         }
1067         return 0;
1068 }
1069
1070 /*
1071  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1072  * whether that can be merged with its predecessor or its successor.
1073  * Or both (it neatly fills a hole).
1074  *
1075  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1076  * certain not to be mapped by the time vma_merge is called; but when
1077  * called for mprotect, it is certain to be already mapped (either at
1078  * an offset within prev, or at the start of next), and the flags of
1079  * this area are about to be changed to vm_flags - and the no-change
1080  * case has already been eliminated.
1081  *
1082  * The following mprotect cases have to be considered, where AAAA is
1083  * the area passed down from mprotect_fixup, never extending beyond one
1084  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1085  *
1086  *     AAAA             AAAA                   AAAA
1087  *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPNNNNNN
1088  *    cannot merge    might become       might become
1089  *                    PPNNNNNNNNNN       PPPPPPPPPPNN
1090  *    mmap, brk or    case 4 below       case 5 below
1091  *    mremap move:
1092  *                        AAAA               AAAA
1093  *                    PPPP    NNNN       PPPPNNNNXXXX
1094  *                    might become       might become
1095  *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
1096  *                    PPPPPPPPNNNN 2 or  PPPPPPPPXXXX 7 or
1097  *                    PPPPNNNNNNNN 3     PPPPXXXXXXXX 8
1098  *
1099  * It is important for case 8 that the vma NNNN overlapping the
1100  * region AAAA is never going to extended over XXXX. Instead XXXX must
1101  * be extended in region AAAA and NNNN must be removed. This way in
1102  * all cases where vma_merge succeeds, the moment vma_adjust drops the
1103  * rmap_locks, the properties of the merged vma will be already
1104  * correct for the whole merged range. Some of those properties like
1105  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1106  * be correct for the whole merged range immediately after the
1107  * rmap_locks are released. Otherwise if XXXX would be removed and
1108  * NNNN would be extended over the XXXX range, remove_migration_ptes
1109  * or other rmap walkers (if working on addresses beyond the "end"
1110  * parameter) may establish ptes with the wrong permissions of NNNN
1111  * instead of the right permissions of XXXX.
1112  */
1113 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1114                         struct vm_area_struct *prev, unsigned long addr,
1115                         unsigned long end, unsigned long vm_flags,
1116                         struct anon_vma *anon_vma, struct file *file,
1117                         pgoff_t pgoff, struct mempolicy *policy,
1118                         struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1119 {
1120         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1121         struct vm_area_struct *area, *next;
1122         int err;
1123
1124         /*
1125          * We later require that vma->vm_flags == vm_flags,
1126          * so this tests vma->vm_flags & VM_SPECIAL, too.
1127          */
1128         if (vm_flags & VM_SPECIAL)
1129                 return NULL;
1130
1131         if (prev)
1132                 next = prev->vm_next;
1133         else
1134                 next = mm->mmap;
1135         area = next;
1136         if (area && area->vm_end == end)                /* cases 6, 7, 8 */
1137                 next = next->vm_next;
1138
1139         /* verify some invariant that must be enforced by the caller */
1140         VM_WARN_ON(prev && addr <= prev->vm_start);
1141         VM_WARN_ON(area && end > area->vm_end);
1142         VM_WARN_ON(addr >= end);
1143
1144         /*
1145          * Can it merge with the predecessor?
1146          */
1147         if (prev && prev->vm_end == addr &&
1148                         mpol_equal(vma_policy(prev), policy) &&
1149                         can_vma_merge_after(prev, vm_flags,
1150                                             anon_vma, file, pgoff,
1151                                             vm_userfaultfd_ctx)) {
1152                 /*
1153                  * OK, it can.  Can we now merge in the successor as well?
1154                  */
1155                 if (next && end == next->vm_start &&
1156                                 mpol_equal(policy, vma_policy(next)) &&
1157                                 can_vma_merge_before(next, vm_flags,
1158                                                      anon_vma, file,
1159                                                      pgoff+pglen,
1160                                                      vm_userfaultfd_ctx) &&
1161                                 is_mergeable_anon_vma(prev->anon_vma,
1162                                                       next->anon_vma, NULL)) {
1163                                                         /* cases 1, 6 */
1164                         err = __vma_adjust(prev, prev->vm_start,
1165                                          next->vm_end, prev->vm_pgoff, NULL,
1166                                          prev);
1167                 } else                                  /* cases 2, 5, 7 */
1168                         err = __vma_adjust(prev, prev->vm_start,
1169                                          end, prev->vm_pgoff, NULL, prev);
1170                 if (err)
1171                         return NULL;
1172                 khugepaged_enter_vma_merge(prev, vm_flags);
1173                 return prev;
1174         }
1175
1176         /*
1177          * Can this new request be merged in front of next?
1178          */
1179         if (next && end == next->vm_start &&
1180                         mpol_equal(policy, vma_policy(next)) &&
1181                         can_vma_merge_before(next, vm_flags,
1182                                              anon_vma, file, pgoff+pglen,
1183                                              vm_userfaultfd_ctx)) {
1184                 if (prev && addr < prev->vm_end)        /* case 4 */
1185                         err = __vma_adjust(prev, prev->vm_start,
1186                                          addr, prev->vm_pgoff, NULL, next);
1187                 else {                                  /* cases 3, 8 */
1188                         err = __vma_adjust(area, addr, next->vm_end,
1189                                          next->vm_pgoff - pglen, NULL, next);
1190                         /*
1191                          * In case 3 area is already equal to next and
1192                          * this is a noop, but in case 8 "area" has
1193                          * been removed and next was expanded over it.
1194                          */
1195                         area = next;
1196                 }
1197                 if (err)
1198                         return NULL;
1199                 khugepaged_enter_vma_merge(area, vm_flags);
1200                 return area;
1201         }
1202
1203         return NULL;
1204 }
1205
1206 /*
1207  * Rough compatbility check to quickly see if it's even worth looking
1208  * at sharing an anon_vma.
1209  *
1210  * They need to have the same vm_file, and the flags can only differ
1211  * in things that mprotect may change.
1212  *
1213  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1214  * we can merge the two vma's. For example, we refuse to merge a vma if
1215  * there is a vm_ops->close() function, because that indicates that the
1216  * driver is doing some kind of reference counting. But that doesn't
1217  * really matter for the anon_vma sharing case.
1218  */
1219 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1220 {
1221         return a->vm_end == b->vm_start &&
1222                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1223                 a->vm_file == b->vm_file &&
1224                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1225                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1226 }
1227
1228 /*
1229  * Do some basic sanity checking to see if we can re-use the anon_vma
1230  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1231  * the same as 'old', the other will be the new one that is trying
1232  * to share the anon_vma.
1233  *
1234  * NOTE! This runs with mm_sem held for reading, so it is possible that
1235  * the anon_vma of 'old' is concurrently in the process of being set up
1236  * by another page fault trying to merge _that_. But that's ok: if it
1237  * is being set up, that automatically means that it will be a singleton
1238  * acceptable for merging, so we can do all of this optimistically. But
1239  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1240  *
1241  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1242  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1243  * is to return an anon_vma that is "complex" due to having gone through
1244  * a fork).
1245  *
1246  * We also make sure that the two vma's are compatible (adjacent,
1247  * and with the same memory policies). That's all stable, even with just
1248  * a read lock on the mm_sem.
1249  */
1250 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1251 {
1252         if (anon_vma_compatible(a, b)) {
1253                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1254
1255                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1256                         return anon_vma;
1257         }
1258         return NULL;
1259 }
1260
1261 /*
1262  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1263  * neighbouring vmas for a suitable anon_vma, before it goes off
1264  * to allocate a new anon_vma.  It checks because a repetitive
1265  * sequence of mprotects and faults may otherwise lead to distinct
1266  * anon_vmas being allocated, preventing vma merge in subsequent
1267  * mprotect.
1268  */
1269 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1270 {
1271         struct anon_vma *anon_vma = NULL;
1272
1273         /* Try next first. */
1274         if (vma->vm_next) {
1275                 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1276                 if (anon_vma)
1277                         return anon_vma;
1278         }
1279
1280         /* Try prev next. */
1281         if (vma->vm_prev)
1282                 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1283
1284         /*
1285          * We might reach here with anon_vma == NULL if we can't find
1286          * any reusable anon_vma.
1287          * There's no absolute need to look only at touching neighbours:
1288          * we could search further afield for "compatible" anon_vmas.
1289          * But it would probably just be a waste of time searching,
1290          * or lead to too many vmas hanging off the same anon_vma.
1291          * We're trying to allow mprotect remerging later on,
1292          * not trying to minimize memory used for anon_vmas.
1293          */
1294         return anon_vma;
1295 }
1296
1297 /*
1298  * If a hint addr is less than mmap_min_addr change hint to be as
1299  * low as possible but still greater than mmap_min_addr
1300  */
1301 static inline unsigned long round_hint_to_min(unsigned long hint)
1302 {
1303         hint &= PAGE_MASK;
1304         if (((void *)hint != NULL) &&
1305             (hint < mmap_min_addr))
1306                 return PAGE_ALIGN(mmap_min_addr);
1307         return hint;
1308 }
1309
1310 static inline int mlock_future_check(struct mm_struct *mm,
1311                                      unsigned long flags,
1312                                      unsigned long len)
1313 {
1314         unsigned long locked, lock_limit;
1315
1316         /*  mlock MCL_FUTURE? */
1317         if (flags & VM_LOCKED) {
1318                 locked = len >> PAGE_SHIFT;
1319                 locked += mm->locked_vm;
1320                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1321                 lock_limit >>= PAGE_SHIFT;
1322                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1323                         return -EAGAIN;
1324         }
1325         return 0;
1326 }
1327
1328 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1329 {
1330         if (S_ISREG(inode->i_mode))
1331                 return MAX_LFS_FILESIZE;
1332
1333         if (S_ISBLK(inode->i_mode))
1334                 return MAX_LFS_FILESIZE;
1335
1336         if (S_ISSOCK(inode->i_mode))
1337                 return MAX_LFS_FILESIZE;
1338
1339         /* Special "we do even unsigned file positions" case */
1340         if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1341                 return 0;
1342
1343         /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1344         return ULONG_MAX;
1345 }
1346
1347 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1348                                 unsigned long pgoff, unsigned long len)
1349 {
1350         u64 maxsize = file_mmap_size_max(file, inode);
1351
1352         if (maxsize && len > maxsize)
1353                 return false;
1354         maxsize -= len;
1355         if (pgoff > maxsize >> PAGE_SHIFT)
1356                 return false;
1357         return true;
1358 }
1359
1360 /*
1361  * The caller must hold down_write(&current->mm->mmap_sem).
1362  */
1363 unsigned long do_mmap(struct file *file, unsigned long addr,
1364                         unsigned long len, unsigned long prot,
1365                         unsigned long flags, vm_flags_t vm_flags,
1366                         unsigned long pgoff, unsigned long *populate,
1367                         struct list_head *uf)
1368 {
1369         struct mm_struct *mm = current->mm;
1370         int pkey = 0;
1371
1372         *populate = 0;
1373
1374         if (!len)
1375                 return -EINVAL;
1376
1377         /*
1378          * Does the application expect PROT_READ to imply PROT_EXEC?
1379          *
1380          * (the exception is when the underlying filesystem is noexec
1381          *  mounted, in which case we dont add PROT_EXEC.)
1382          */
1383         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1384                 if (!(file && path_noexec(&file->f_path)))
1385                         prot |= PROT_EXEC;
1386
1387         /* force arch specific MAP_FIXED handling in get_unmapped_area */
1388         if (flags & MAP_FIXED_NOREPLACE)
1389                 flags |= MAP_FIXED;
1390
1391         if (!(flags & MAP_FIXED))
1392                 addr = round_hint_to_min(addr);
1393
1394         /* Careful about overflows.. */
1395         len = PAGE_ALIGN(len);
1396         if (!len)
1397                 return -ENOMEM;
1398
1399         /* offset overflow? */
1400         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1401                 return -EOVERFLOW;
1402
1403         /* Too many mappings? */
1404         if (mm->map_count > sysctl_max_map_count)
1405                 return -ENOMEM;
1406
1407         /* Obtain the address to map to. we verify (or select) it and ensure
1408          * that it represents a valid section of the address space.
1409          */
1410         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1411         if (IS_ERR_VALUE(addr))
1412                 return addr;
1413
1414         if (flags & MAP_FIXED_NOREPLACE) {
1415                 struct vm_area_struct *vma = find_vma(mm, addr);
1416
1417                 if (vma && vma->vm_start < addr + len)
1418                         return -EEXIST;
1419         }
1420
1421         if (prot == PROT_EXEC) {
1422                 pkey = execute_only_pkey(mm);
1423                 if (pkey < 0)
1424                         pkey = 0;
1425         }
1426
1427         /* Do simple checking here so the lower-level routines won't have
1428          * to. we assume access permissions have been handled by the open
1429          * of the memory object, so we don't do any here.
1430          */
1431         vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1432                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1433
1434         if (flags & MAP_LOCKED)
1435                 if (!can_do_mlock())
1436                         return -EPERM;
1437
1438         if (mlock_future_check(mm, vm_flags, len))
1439                 return -EAGAIN;
1440
1441         if (file) {
1442                 struct inode *inode = file_inode(file);
1443                 unsigned long flags_mask;
1444
1445                 if (!file_mmap_ok(file, inode, pgoff, len))
1446                         return -EOVERFLOW;
1447
1448                 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1449
1450                 switch (flags & MAP_TYPE) {
1451                 case MAP_SHARED:
1452                         /*
1453                          * Force use of MAP_SHARED_VALIDATE with non-legacy
1454                          * flags. E.g. MAP_SYNC is dangerous to use with
1455                          * MAP_SHARED as you don't know which consistency model
1456                          * you will get. We silently ignore unsupported flags
1457                          * with MAP_SHARED to preserve backward compatibility.
1458                          */
1459                         flags &= LEGACY_MAP_MASK;
1460                         /* fall through */
1461                 case MAP_SHARED_VALIDATE:
1462                         if (flags & ~flags_mask)
1463                                 return -EOPNOTSUPP;
1464                         if (prot & PROT_WRITE) {
1465                                 if (!(file->f_mode & FMODE_WRITE))
1466                                         return -EACCES;
1467                                 if (IS_SWAPFILE(file->f_mapping->host))
1468                                         return -ETXTBSY;
1469                         }
1470
1471                         /*
1472                          * Make sure we don't allow writing to an append-only
1473                          * file..
1474                          */
1475                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1476                                 return -EACCES;
1477
1478                         /*
1479                          * Make sure there are no mandatory locks on the file.
1480                          */
1481                         if (locks_verify_locked(file))
1482                                 return -EAGAIN;
1483
1484                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1485                         if (!(file->f_mode & FMODE_WRITE))
1486                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1487
1488                         /* fall through */
1489                 case MAP_PRIVATE:
1490                         if (!(file->f_mode & FMODE_READ))
1491                                 return -EACCES;
1492                         if (path_noexec(&file->f_path)) {
1493                                 if (vm_flags & VM_EXEC)
1494                                         return -EPERM;
1495                                 vm_flags &= ~VM_MAYEXEC;
1496                         }
1497
1498                         if (!file->f_op->mmap)
1499                                 return -ENODEV;
1500                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1501                                 return -EINVAL;
1502                         break;
1503
1504                 default:
1505                         return -EINVAL;
1506                 }
1507         } else {
1508                 switch (flags & MAP_TYPE) {
1509                 case MAP_SHARED:
1510                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1511                                 return -EINVAL;
1512                         /*
1513                          * Ignore pgoff.
1514                          */
1515                         pgoff = 0;
1516                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1517                         break;
1518                 case MAP_PRIVATE:
1519                         /*
1520                          * Set pgoff according to addr for anon_vma.
1521                          */
1522                         pgoff = addr >> PAGE_SHIFT;
1523                         break;
1524                 default:
1525                         return -EINVAL;
1526                 }
1527         }
1528
1529         /*
1530          * Set 'VM_NORESERVE' if we should not account for the
1531          * memory use of this mapping.
1532          */
1533         if (flags & MAP_NORESERVE) {
1534                 /* We honor MAP_NORESERVE if allowed to overcommit */
1535                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1536                         vm_flags |= VM_NORESERVE;
1537
1538                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1539                 if (file && is_file_hugepages(file))
1540                         vm_flags |= VM_NORESERVE;
1541         }
1542
1543         addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1544         if (!IS_ERR_VALUE(addr) &&
1545             ((vm_flags & VM_LOCKED) ||
1546              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1547                 *populate = len;
1548         return addr;
1549 }
1550
1551 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1552                               unsigned long prot, unsigned long flags,
1553                               unsigned long fd, unsigned long pgoff)
1554 {
1555         struct file *file = NULL;
1556         unsigned long retval;
1557
1558         if (!(flags & MAP_ANONYMOUS)) {
1559                 audit_mmap_fd(fd, flags);
1560                 file = fget(fd);
1561                 if (!file)
1562                         return -EBADF;
1563                 if (is_file_hugepages(file))
1564                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1565                 retval = -EINVAL;
1566                 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1567                         goto out_fput;
1568         } else if (flags & MAP_HUGETLB) {
1569                 struct user_struct *user = NULL;
1570                 struct hstate *hs;
1571
1572                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1573                 if (!hs)
1574                         return -EINVAL;
1575
1576                 len = ALIGN(len, huge_page_size(hs));
1577                 /*
1578                  * VM_NORESERVE is used because the reservations will be
1579                  * taken when vm_ops->mmap() is called
1580                  * A dummy user value is used because we are not locking
1581                  * memory so no accounting is necessary
1582                  */
1583                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1584                                 VM_NORESERVE,
1585                                 &user, HUGETLB_ANONHUGE_INODE,
1586                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1587                 if (IS_ERR(file))
1588                         return PTR_ERR(file);
1589         }
1590
1591         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1592
1593         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1594 out_fput:
1595         if (file)
1596                 fput(file);
1597         return retval;
1598 }
1599
1600 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1601                 unsigned long, prot, unsigned long, flags,
1602                 unsigned long, fd, unsigned long, pgoff)
1603 {
1604         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1605 }
1606
1607 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1608 struct mmap_arg_struct {
1609         unsigned long addr;
1610         unsigned long len;
1611         unsigned long prot;
1612         unsigned long flags;
1613         unsigned long fd;
1614         unsigned long offset;
1615 };
1616
1617 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1618 {
1619         struct mmap_arg_struct a;
1620
1621         if (copy_from_user(&a, arg, sizeof(a)))
1622                 return -EFAULT;
1623         if (offset_in_page(a.offset))
1624                 return -EINVAL;
1625
1626         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1627                                a.offset >> PAGE_SHIFT);
1628 }
1629 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1630
1631 /*
1632  * Some shared mappings will want the pages marked read-only
1633  * to track write events. If so, we'll downgrade vm_page_prot
1634  * to the private version (using protection_map[] without the
1635  * VM_SHARED bit).
1636  */
1637 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1638 {
1639         vm_flags_t vm_flags = vma->vm_flags;
1640         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1641
1642         /* If it was private or non-writable, the write bit is already clear */
1643         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1644                 return 0;
1645
1646         /* The backer wishes to know when pages are first written to? */
1647         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1648                 return 1;
1649
1650         /* The open routine did something to the protections that pgprot_modify
1651          * won't preserve? */
1652         if (pgprot_val(vm_page_prot) !=
1653             pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1654                 return 0;
1655
1656         /* Do we need to track softdirty? */
1657         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1658                 return 1;
1659
1660         /* Specialty mapping? */
1661         if (vm_flags & VM_PFNMAP)
1662                 return 0;
1663
1664         /* Can the mapping track the dirty pages? */
1665         return vma->vm_file && vma->vm_file->f_mapping &&
1666                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1667 }
1668
1669 /*
1670  * We account for memory if it's a private writeable mapping,
1671  * not hugepages and VM_NORESERVE wasn't set.
1672  */
1673 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1674 {
1675         /*
1676          * hugetlb has its own accounting separate from the core VM
1677          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1678          */
1679         if (file && is_file_hugepages(file))
1680                 return 0;
1681
1682         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1683 }
1684
1685 unsigned long mmap_region(struct file *file, unsigned long addr,
1686                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1687                 struct list_head *uf)
1688 {
1689         struct mm_struct *mm = current->mm;
1690         struct vm_area_struct *vma, *prev;
1691         int error;
1692         struct rb_node **rb_link, *rb_parent;
1693         unsigned long charged = 0;
1694
1695         /* Check against address space limit. */
1696         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1697                 unsigned long nr_pages;
1698
1699                 /*
1700                  * MAP_FIXED may remove pages of mappings that intersects with
1701                  * requested mapping. Account for the pages it would unmap.
1702                  */
1703                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1704
1705                 if (!may_expand_vm(mm, vm_flags,
1706                                         (len >> PAGE_SHIFT) - nr_pages))
1707                         return -ENOMEM;
1708         }
1709
1710         /* Clear old maps */
1711         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1712                               &rb_parent)) {
1713                 if (do_munmap(mm, addr, len, uf))
1714                         return -ENOMEM;
1715         }
1716
1717         /*
1718          * Private writable mapping: check memory availability
1719          */
1720         if (accountable_mapping(file, vm_flags)) {
1721                 charged = len >> PAGE_SHIFT;
1722                 if (security_vm_enough_memory_mm(mm, charged))
1723                         return -ENOMEM;
1724                 vm_flags |= VM_ACCOUNT;
1725         }
1726
1727         /*
1728          * Can we just expand an old mapping?
1729          */
1730         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1731                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1732         if (vma)
1733                 goto out;
1734
1735         /*
1736          * Determine the object being mapped and call the appropriate
1737          * specific mapper. the address has already been validated, but
1738          * not unmapped, but the maps are removed from the list.
1739          */
1740         vma = vm_area_alloc(mm);
1741         if (!vma) {
1742                 error = -ENOMEM;
1743                 goto unacct_error;
1744         }
1745
1746         vma->vm_start = addr;
1747         vma->vm_end = addr + len;
1748         vma->vm_flags = vm_flags;
1749         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1750         vma->vm_pgoff = pgoff;
1751
1752         if (file) {
1753                 if (vm_flags & VM_DENYWRITE) {
1754                         error = deny_write_access(file);
1755                         if (error)
1756                                 goto free_vma;
1757                 }
1758                 if (vm_flags & VM_SHARED) {
1759                         error = mapping_map_writable(file->f_mapping);
1760                         if (error)
1761                                 goto allow_write_and_free_vma;
1762                 }
1763
1764                 /* ->mmap() can change vma->vm_file, but must guarantee that
1765                  * vma_link() below can deny write-access if VM_DENYWRITE is set
1766                  * and map writably if VM_SHARED is set. This usually means the
1767                  * new file must not have been exposed to user-space, yet.
1768                  */
1769                 vma->vm_file = get_file(file);
1770                 error = call_mmap(file, vma);
1771                 if (error)
1772                         goto unmap_and_free_vma;
1773
1774                 /* Can addr have changed??
1775                  *
1776                  * Answer: Yes, several device drivers can do it in their
1777                  *         f_op->mmap method. -DaveM
1778                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1779                  *      be updated for vma_link()
1780                  */
1781                 WARN_ON_ONCE(addr != vma->vm_start);
1782
1783                 addr = vma->vm_start;
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         vma_link(mm, vma, prev, rb_link, rb_parent);
1794         /* Once vma denies write, undo our temporary denial count */
1795         if (file) {
1796                 if (vm_flags & VM_SHARED)
1797                         mapping_unmap_writable(file->f_mapping);
1798                 if (vm_flags & VM_DENYWRITE)
1799                         allow_write_access(file);
1800         }
1801         file = vma->vm_file;
1802 out:
1803         perf_event_mmap(vma);
1804
1805         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1806         if (vm_flags & VM_LOCKED) {
1807                 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1808                                         is_vm_hugetlb_page(vma) ||
1809                                         vma == get_gate_vma(current->mm))
1810                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1811                 else
1812                         mm->locked_vm += (len >> PAGE_SHIFT);
1813         }
1814
1815         if (file)
1816                 uprobe_mmap(vma);
1817
1818         /*
1819          * New (or expanded) vma always get soft dirty status.
1820          * Otherwise user-space soft-dirty page tracker won't
1821          * be able to distinguish situation when vma area unmapped,
1822          * then new mapped in-place (which must be aimed as
1823          * a completely new data area).
1824          */
1825         vma->vm_flags |= VM_SOFTDIRTY;
1826
1827         vma_set_page_prot(vma);
1828
1829         return addr;
1830
1831 unmap_and_free_vma:
1832         vma->vm_file = NULL;
1833         fput(file);
1834
1835         /* Undo any partial mapping done by a device driver. */
1836         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1837         charged = 0;
1838         if (vm_flags & VM_SHARED)
1839                 mapping_unmap_writable(file->f_mapping);
1840 allow_write_and_free_vma:
1841         if (vm_flags & VM_DENYWRITE)
1842                 allow_write_access(file);
1843 free_vma:
1844         vm_area_free(vma);
1845 unacct_error:
1846         if (charged)
1847                 vm_unacct_memory(charged);
1848         return error;
1849 }
1850
1851 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1852 {
1853         /*
1854          * We implement the search by looking for an rbtree node that
1855          * immediately follows a suitable gap. That is,
1856          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1857          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1858          * - gap_end - gap_start >= length
1859          */
1860
1861         struct mm_struct *mm = current->mm;
1862         struct vm_area_struct *vma;
1863         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1864
1865         /* Adjust search length to account for worst case alignment overhead */
1866         length = info->length + info->align_mask;
1867         if (length < info->length)
1868                 return -ENOMEM;
1869
1870         /* Adjust search limits by the desired length */
1871         if (info->high_limit < length)
1872                 return -ENOMEM;
1873         high_limit = info->high_limit - length;
1874
1875         if (info->low_limit > high_limit)
1876                 return -ENOMEM;
1877         low_limit = info->low_limit + length;
1878
1879         /* Check if rbtree root looks promising */
1880         if (RB_EMPTY_ROOT(&mm->mm_rb))
1881                 goto check_highest;
1882         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1883         if (vma->rb_subtree_gap < length)
1884                 goto check_highest;
1885
1886         while (true) {
1887                 /* Visit left subtree if it looks promising */
1888                 gap_end = vm_start_gap(vma);
1889                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1890                         struct vm_area_struct *left =
1891                                 rb_entry(vma->vm_rb.rb_left,
1892                                          struct vm_area_struct, vm_rb);
1893                         if (left->rb_subtree_gap >= length) {
1894                                 vma = left;
1895                                 continue;
1896                         }
1897                 }
1898
1899                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1900 check_current:
1901                 /* Check if current node has a suitable gap */
1902                 if (gap_start > high_limit)
1903                         return -ENOMEM;
1904                 if (gap_end >= low_limit &&
1905                     gap_end > gap_start && gap_end - gap_start >= length)
1906                         goto found;
1907
1908                 /* Visit right subtree if it looks promising */
1909                 if (vma->vm_rb.rb_right) {
1910                         struct vm_area_struct *right =
1911                                 rb_entry(vma->vm_rb.rb_right,
1912                                          struct vm_area_struct, vm_rb);
1913                         if (right->rb_subtree_gap >= length) {
1914                                 vma = right;
1915                                 continue;
1916                         }
1917                 }
1918
1919                 /* Go back up the rbtree to find next candidate node */
1920                 while (true) {
1921                         struct rb_node *prev = &vma->vm_rb;
1922                         if (!rb_parent(prev))
1923                                 goto check_highest;
1924                         vma = rb_entry(rb_parent(prev),
1925                                        struct vm_area_struct, vm_rb);
1926                         if (prev == vma->vm_rb.rb_left) {
1927                                 gap_start = vm_end_gap(vma->vm_prev);
1928                                 gap_end = vm_start_gap(vma);
1929                                 goto check_current;
1930                         }
1931                 }
1932         }
1933
1934 check_highest:
1935         /* Check highest gap, which does not precede any rbtree node */
1936         gap_start = mm->highest_vm_end;
1937         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1938         if (gap_start > high_limit)
1939                 return -ENOMEM;
1940
1941 found:
1942         /* We found a suitable gap. Clip it with the original low_limit. */
1943         if (gap_start < info->low_limit)
1944                 gap_start = info->low_limit;
1945
1946         /* Adjust gap address to the desired alignment */
1947         gap_start += (info->align_offset - gap_start) & info->align_mask;
1948
1949         VM_BUG_ON(gap_start + info->length > info->high_limit);
1950         VM_BUG_ON(gap_start + info->length > gap_end);
1951         return gap_start;
1952 }
1953
1954 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1955 {
1956         struct mm_struct *mm = current->mm;
1957         struct vm_area_struct *vma;
1958         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1959
1960         /* Adjust search length to account for worst case alignment overhead */
1961         length = info->length + info->align_mask;
1962         if (length < info->length)
1963                 return -ENOMEM;
1964
1965         /*
1966          * Adjust search limits by the desired length.
1967          * See implementation comment at top of unmapped_area().
1968          */
1969         gap_end = info->high_limit;
1970         if (gap_end < length)
1971                 return -ENOMEM;
1972         high_limit = gap_end - length;
1973
1974         if (info->low_limit > high_limit)
1975                 return -ENOMEM;
1976         low_limit = info->low_limit + length;
1977
1978         /* Check highest gap, which does not precede any rbtree node */
1979         gap_start = mm->highest_vm_end;
1980         if (gap_start <= high_limit)
1981                 goto found_highest;
1982
1983         /* Check if rbtree root looks promising */
1984         if (RB_EMPTY_ROOT(&mm->mm_rb))
1985                 return -ENOMEM;
1986         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1987         if (vma->rb_subtree_gap < length)
1988                 return -ENOMEM;
1989
1990         while (true) {
1991                 /* Visit right subtree if it looks promising */
1992                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1993                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1994                         struct vm_area_struct *right =
1995                                 rb_entry(vma->vm_rb.rb_right,
1996                                          struct vm_area_struct, vm_rb);
1997                         if (right->rb_subtree_gap >= length) {
1998                                 vma = right;
1999                                 continue;
2000                         }
2001                 }
2002
2003 check_current:
2004                 /* Check if current node has a suitable gap */
2005                 gap_end = vm_start_gap(vma);
2006                 if (gap_end < low_limit)
2007                         return -ENOMEM;
2008                 if (gap_start <= high_limit &&
2009                     gap_end > gap_start && gap_end - gap_start >= length)
2010                         goto found;
2011
2012                 /* Visit left subtree if it looks promising */
2013                 if (vma->vm_rb.rb_left) {
2014                         struct vm_area_struct *left =
2015                                 rb_entry(vma->vm_rb.rb_left,
2016                                          struct vm_area_struct, vm_rb);
2017                         if (left->rb_subtree_gap >= length) {
2018                                 vma = left;
2019                                 continue;
2020                         }
2021                 }
2022
2023                 /* Go back up the rbtree to find next candidate node */
2024                 while (true) {
2025                         struct rb_node *prev = &vma->vm_rb;
2026                         if (!rb_parent(prev))
2027                                 return -ENOMEM;
2028                         vma = rb_entry(rb_parent(prev),
2029                                        struct vm_area_struct, vm_rb);
2030                         if (prev == vma->vm_rb.rb_right) {
2031                                 gap_start = vma->vm_prev ?
2032                                         vm_end_gap(vma->vm_prev) : 0;
2033                                 goto check_current;
2034                         }
2035                 }
2036         }
2037
2038 found:
2039         /* We found a suitable gap. Clip it with the original high_limit. */
2040         if (gap_end > info->high_limit)
2041                 gap_end = info->high_limit;
2042
2043 found_highest:
2044         /* Compute highest gap address at the desired alignment */
2045         gap_end -= info->length;
2046         gap_end -= (gap_end - info->align_offset) & info->align_mask;
2047
2048         VM_BUG_ON(gap_end < info->low_limit);
2049         VM_BUG_ON(gap_end < gap_start);
2050         return gap_end;
2051 }
2052
2053
2054 #ifndef arch_get_mmap_end
2055 #define arch_get_mmap_end(addr) (TASK_SIZE)
2056 #endif
2057
2058 #ifndef arch_get_mmap_base
2059 #define arch_get_mmap_base(addr, base) (base)
2060 #endif
2061
2062 /* Get an address range which is currently unmapped.
2063  * For shmat() with addr=0.
2064  *
2065  * Ugly calling convention alert:
2066  * Return value with the low bits set means error value,
2067  * ie
2068  *      if (ret & ~PAGE_MASK)
2069  *              error = ret;
2070  *
2071  * This function "knows" that -ENOMEM has the bits set.
2072  */
2073 #ifndef HAVE_ARCH_UNMAPPED_AREA
2074 unsigned long
2075 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2076                 unsigned long len, unsigned long pgoff, unsigned long flags)
2077 {
2078         struct mm_struct *mm = current->mm;
2079         struct vm_area_struct *vma, *prev;
2080         struct vm_unmapped_area_info info;
2081         const unsigned long mmap_end = arch_get_mmap_end(addr);
2082
2083         if (len > mmap_end - mmap_min_addr)
2084                 return -ENOMEM;
2085
2086         if (flags & MAP_FIXED)
2087                 return addr;
2088
2089         if (addr) {
2090                 addr = PAGE_ALIGN(addr);
2091                 vma = find_vma_prev(mm, addr, &prev);
2092                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2093                     (!vma || addr + len <= vm_start_gap(vma)) &&
2094                     (!prev || addr >= vm_end_gap(prev)))
2095                         return addr;
2096         }
2097
2098         info.flags = 0;
2099         info.length = len;
2100         info.low_limit = mm->mmap_base;
2101         info.high_limit = mmap_end;
2102         info.align_mask = 0;
2103         return vm_unmapped_area(&info);
2104 }
2105 #endif
2106
2107 /*
2108  * This mmap-allocator allocates new areas top-down from below the
2109  * stack's low limit (the base):
2110  */
2111 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2112 unsigned long
2113 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2114                           unsigned long len, unsigned long pgoff,
2115                           unsigned long flags)
2116 {
2117         struct vm_area_struct *vma, *prev;
2118         struct mm_struct *mm = current->mm;
2119         struct vm_unmapped_area_info info;
2120         const unsigned long mmap_end = arch_get_mmap_end(addr);
2121
2122         /* requested length too big for entire address space */
2123         if (len > mmap_end - mmap_min_addr)
2124                 return -ENOMEM;
2125
2126         if (flags & MAP_FIXED)
2127                 return addr;
2128
2129         /* requesting a specific address */
2130         if (addr) {
2131                 addr = PAGE_ALIGN(addr);
2132                 vma = find_vma_prev(mm, addr, &prev);
2133                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2134                                 (!vma || addr + len <= vm_start_gap(vma)) &&
2135                                 (!prev || addr >= vm_end_gap(prev)))
2136                         return addr;
2137         }
2138
2139         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2140         info.length = len;
2141         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2142         info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2143         info.align_mask = 0;
2144         addr = vm_unmapped_area(&info);
2145
2146         /*
2147          * A failed mmap() very likely causes application failure,
2148          * so fall back to the bottom-up function here. This scenario
2149          * can happen with large stack limits and large mmap()
2150          * allocations.
2151          */
2152         if (offset_in_page(addr)) {
2153                 VM_BUG_ON(addr != -ENOMEM);
2154                 info.flags = 0;
2155                 info.low_limit = TASK_UNMAPPED_BASE;
2156                 info.high_limit = mmap_end;
2157                 addr = vm_unmapped_area(&info);
2158         }
2159
2160         return addr;
2161 }
2162 #endif
2163
2164 unsigned long
2165 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2166                 unsigned long pgoff, unsigned long flags)
2167 {
2168         unsigned long (*get_area)(struct file *, unsigned long,
2169                                   unsigned long, unsigned long, unsigned long);
2170
2171         unsigned long error = arch_mmap_check(addr, len, flags);
2172         if (error)
2173                 return error;
2174
2175         /* Careful about overflows.. */
2176         if (len > TASK_SIZE)
2177                 return -ENOMEM;
2178
2179         get_area = current->mm->get_unmapped_area;
2180         if (file) {
2181                 if (file->f_op->get_unmapped_area)
2182                         get_area = file->f_op->get_unmapped_area;
2183         } else if (flags & MAP_SHARED) {
2184                 /*
2185                  * mmap_region() will call shmem_zero_setup() to create a file,
2186                  * so use shmem's get_unmapped_area in case it can be huge.
2187                  * do_mmap_pgoff() will clear pgoff, so match alignment.
2188                  */
2189                 pgoff = 0;
2190                 get_area = shmem_get_unmapped_area;
2191         }
2192
2193         addr = get_area(file, addr, len, pgoff, flags);
2194         if (IS_ERR_VALUE(addr))
2195                 return addr;
2196
2197         if (addr > TASK_SIZE - len)
2198                 return -ENOMEM;
2199         if (offset_in_page(addr))
2200                 return -EINVAL;
2201
2202         error = security_mmap_addr(addr);
2203         return error ? error : addr;
2204 }
2205
2206 EXPORT_SYMBOL(get_unmapped_area);
2207
2208 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2209 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2210 {
2211         struct rb_node *rb_node;
2212         struct vm_area_struct *vma;
2213
2214         /* Check the cache first. */
2215         vma = vmacache_find(mm, addr);
2216         if (likely(vma))
2217                 return vma;
2218
2219         rb_node = mm->mm_rb.rb_node;
2220
2221         while (rb_node) {
2222                 struct vm_area_struct *tmp;
2223
2224                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2225
2226                 if (tmp->vm_end > addr) {
2227                         vma = tmp;
2228                         if (tmp->vm_start <= addr)
2229                                 break;
2230                         rb_node = rb_node->rb_left;
2231                 } else
2232                         rb_node = rb_node->rb_right;
2233         }
2234
2235         if (vma)
2236                 vmacache_update(addr, vma);
2237         return vma;
2238 }
2239
2240 EXPORT_SYMBOL(find_vma);
2241
2242 /*
2243  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2244  */
2245 struct vm_area_struct *
2246 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2247                         struct vm_area_struct **pprev)
2248 {
2249         struct vm_area_struct *vma;
2250
2251         vma = find_vma(mm, addr);
2252         if (vma) {
2253                 *pprev = vma->vm_prev;
2254         } else {
2255                 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2256
2257                 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2258         }
2259         return vma;
2260 }
2261
2262 /*
2263  * Verify that the stack growth is acceptable and
2264  * update accounting. This is shared with both the
2265  * grow-up and grow-down cases.
2266  */
2267 static int acct_stack_growth(struct vm_area_struct *vma,
2268                              unsigned long size, unsigned long grow)
2269 {
2270         struct mm_struct *mm = vma->vm_mm;
2271         unsigned long new_start;
2272
2273         /* address space limit tests */
2274         if (!may_expand_vm(mm, vma->vm_flags, grow))
2275                 return -ENOMEM;
2276
2277         /* Stack limit test */
2278         if (size > rlimit(RLIMIT_STACK))
2279                 return -ENOMEM;
2280
2281         /* mlock limit tests */
2282         if (vma->vm_flags & VM_LOCKED) {
2283                 unsigned long locked;
2284                 unsigned long limit;
2285                 locked = mm->locked_vm + grow;
2286                 limit = rlimit(RLIMIT_MEMLOCK);
2287                 limit >>= PAGE_SHIFT;
2288                 if (locked > limit && !capable(CAP_IPC_LOCK))
2289                         return -ENOMEM;
2290         }
2291
2292         /* Check to ensure the stack will not grow into a hugetlb-only region */
2293         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2294                         vma->vm_end - size;
2295         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2296                 return -EFAULT;
2297
2298         /*
2299          * Overcommit..  This must be the final test, as it will
2300          * update security statistics.
2301          */
2302         if (security_vm_enough_memory_mm(mm, grow))
2303                 return -ENOMEM;
2304
2305         return 0;
2306 }
2307
2308 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2309 /*
2310  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2311  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2312  */
2313 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2314 {
2315         struct mm_struct *mm = vma->vm_mm;
2316         struct vm_area_struct *next;
2317         unsigned long gap_addr;
2318         int error = 0;
2319
2320         if (!(vma->vm_flags & VM_GROWSUP))
2321                 return -EFAULT;
2322
2323         /* Guard against exceeding limits of the address space. */
2324         address &= PAGE_MASK;
2325         if (address >= (TASK_SIZE & PAGE_MASK))
2326                 return -ENOMEM;
2327         address += PAGE_SIZE;
2328
2329         /* Enforce stack_guard_gap */
2330         gap_addr = address + stack_guard_gap;
2331
2332         /* Guard against overflow */
2333         if (gap_addr < address || gap_addr > TASK_SIZE)
2334                 gap_addr = TASK_SIZE;
2335
2336         next = vma->vm_next;
2337         if (next && next->vm_start < gap_addr &&
2338                         (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2339                 if (!(next->vm_flags & VM_GROWSUP))
2340                         return -ENOMEM;
2341                 /* Check that both stack segments have the same anon_vma? */
2342         }
2343
2344         /* We must make sure the anon_vma is allocated. */
2345         if (unlikely(anon_vma_prepare(vma)))
2346                 return -ENOMEM;
2347
2348         /*
2349          * vma->vm_start/vm_end cannot change under us because the caller
2350          * is required to hold the mmap_sem in read mode.  We need the
2351          * anon_vma lock to serialize against concurrent expand_stacks.
2352          */
2353         anon_vma_lock_write(vma->anon_vma);
2354
2355         /* Somebody else might have raced and expanded it already */
2356         if (address > vma->vm_end) {
2357                 unsigned long size, grow;
2358
2359                 size = address - vma->vm_start;
2360                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2361
2362                 error = -ENOMEM;
2363                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2364                         error = acct_stack_growth(vma, size, grow);
2365                         if (!error) {
2366                                 /*
2367                                  * vma_gap_update() doesn't support concurrent
2368                                  * updates, but we only hold a shared mmap_sem
2369                                  * lock here, so we need to protect against
2370                                  * concurrent vma expansions.
2371                                  * anon_vma_lock_write() doesn't help here, as
2372                                  * we don't guarantee that all growable vmas
2373                                  * in a mm share the same root anon vma.
2374                                  * So, we reuse mm->page_table_lock to guard
2375                                  * against concurrent vma expansions.
2376                                  */
2377                                 spin_lock(&mm->page_table_lock);
2378                                 if (vma->vm_flags & VM_LOCKED)
2379                                         mm->locked_vm += grow;
2380                                 vm_stat_account(mm, vma->vm_flags, grow);
2381                                 anon_vma_interval_tree_pre_update_vma(vma);
2382                                 vma->vm_end = address;
2383                                 anon_vma_interval_tree_post_update_vma(vma);
2384                                 if (vma->vm_next)
2385                                         vma_gap_update(vma->vm_next);
2386                                 else
2387                                         mm->highest_vm_end = vm_end_gap(vma);
2388                                 spin_unlock(&mm->page_table_lock);
2389
2390                                 perf_event_mmap(vma);
2391                         }
2392                 }
2393         }
2394         anon_vma_unlock_write(vma->anon_vma);
2395         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2396         validate_mm(mm);
2397         return error;
2398 }
2399 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2400
2401 /*
2402  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2403  */
2404 int expand_downwards(struct vm_area_struct *vma,
2405                                    unsigned long address)
2406 {
2407         struct mm_struct *mm = vma->vm_mm;
2408         struct vm_area_struct *prev;
2409         int error = 0;
2410
2411         address &= PAGE_MASK;
2412         if (address < mmap_min_addr)
2413                 return -EPERM;
2414
2415         /* Enforce stack_guard_gap */
2416         prev = vma->vm_prev;
2417         /* Check that both stack segments have the same anon_vma? */
2418         if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2419                         (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2420                 if (address - prev->vm_end < stack_guard_gap)
2421                         return -ENOMEM;
2422         }
2423
2424         /* We must make sure the anon_vma is allocated. */
2425         if (unlikely(anon_vma_prepare(vma)))
2426                 return -ENOMEM;
2427
2428         /*
2429          * vma->vm_start/vm_end cannot change under us because the caller
2430          * is required to hold the mmap_sem in read mode.  We need the
2431          * anon_vma lock to serialize against concurrent expand_stacks.
2432          */
2433         anon_vma_lock_write(vma->anon_vma);
2434
2435         /* Somebody else might have raced and expanded it already */
2436         if (address < vma->vm_start) {
2437                 unsigned long size, grow;
2438
2439                 size = vma->vm_end - address;
2440                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2441
2442                 error = -ENOMEM;
2443                 if (grow <= vma->vm_pgoff) {
2444                         error = acct_stack_growth(vma, size, grow);
2445                         if (!error) {
2446                                 /*
2447                                  * vma_gap_update() doesn't support concurrent
2448                                  * updates, but we only hold a shared mmap_sem
2449                                  * lock here, so we need to protect against
2450                                  * concurrent vma expansions.
2451                                  * anon_vma_lock_write() doesn't help here, as
2452                                  * we don't guarantee that all growable vmas
2453                                  * in a mm share the same root anon vma.
2454                                  * So, we reuse mm->page_table_lock to guard
2455                                  * against concurrent vma expansions.
2456                                  */
2457                                 spin_lock(&mm->page_table_lock);
2458                                 if (vma->vm_flags & VM_LOCKED)
2459                                         mm->locked_vm += grow;
2460                                 vm_stat_account(mm, vma->vm_flags, grow);
2461                                 anon_vma_interval_tree_pre_update_vma(vma);
2462                                 vma->vm_start = address;
2463                                 vma->vm_pgoff -= grow;
2464                                 anon_vma_interval_tree_post_update_vma(vma);
2465                                 vma_gap_update(vma);
2466                                 spin_unlock(&mm->page_table_lock);
2467
2468                                 perf_event_mmap(vma);
2469                         }
2470                 }
2471         }
2472         anon_vma_unlock_write(vma->anon_vma);
2473         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2474         validate_mm(mm);
2475         return error;
2476 }
2477
2478 /* enforced gap between the expanding stack and other mappings. */
2479 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2480
2481 static int __init cmdline_parse_stack_guard_gap(char *p)
2482 {
2483         unsigned long val;
2484         char *endptr;
2485
2486         val = simple_strtoul(p, &endptr, 10);
2487         if (!*endptr)
2488                 stack_guard_gap = val << PAGE_SHIFT;
2489
2490         return 0;
2491 }
2492 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2493
2494 #ifdef CONFIG_STACK_GROWSUP
2495 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2496 {
2497         return expand_upwards(vma, address);
2498 }
2499
2500 struct vm_area_struct *
2501 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2502 {
2503         struct vm_area_struct *vma, *prev;
2504
2505         addr &= PAGE_MASK;
2506         vma = find_vma_prev(mm, addr, &prev);
2507         if (vma && (vma->vm_start <= addr))
2508                 return vma;
2509         /* don't alter vm_end if the coredump is running */
2510         if (!prev || !mmget_still_valid(mm) || expand_stack(prev, addr))
2511                 return NULL;
2512         if (prev->vm_flags & VM_LOCKED)
2513                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2514         return prev;
2515 }
2516 #else
2517 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2518 {
2519         return expand_downwards(vma, address);
2520 }
2521
2522 struct vm_area_struct *
2523 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2524 {
2525         struct vm_area_struct *vma;
2526         unsigned long start;
2527
2528         addr &= PAGE_MASK;
2529         vma = find_vma(mm, addr);
2530         if (!vma)
2531                 return NULL;
2532         if (vma->vm_start <= addr)
2533                 return vma;
2534         if (!(vma->vm_flags & VM_GROWSDOWN))
2535                 return NULL;
2536         /* don't alter vm_start if the coredump is running */
2537         if (!mmget_still_valid(mm))
2538                 return NULL;
2539         start = vma->vm_start;
2540         if (expand_stack(vma, addr))
2541                 return NULL;
2542         if (vma->vm_flags & VM_LOCKED)
2543                 populate_vma_page_range(vma, addr, start, NULL);
2544         return vma;
2545 }
2546 #endif
2547
2548 EXPORT_SYMBOL_GPL(find_extend_vma);
2549
2550 /*
2551  * Ok - we have the memory areas we should free on the vma list,
2552  * so release them, and do the vma updates.
2553  *
2554  * Called with the mm semaphore held.
2555  */
2556 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2557 {
2558         unsigned long nr_accounted = 0;
2559
2560         /* Update high watermark before we lower total_vm */
2561         update_hiwater_vm(mm);
2562         do {
2563                 long nrpages = vma_pages(vma);
2564
2565                 if (vma->vm_flags & VM_ACCOUNT)
2566                         nr_accounted += nrpages;
2567                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2568                 vma = remove_vma(vma);
2569         } while (vma);
2570         vm_unacct_memory(nr_accounted);
2571         validate_mm(mm);
2572 }
2573
2574 /*
2575  * Get rid of page table information in the indicated region.
2576  *
2577  * Called with the mm semaphore held.
2578  */
2579 static void unmap_region(struct mm_struct *mm,
2580                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2581                 unsigned long start, unsigned long end)
2582 {
2583         struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2584         struct mmu_gather tlb;
2585
2586         lru_add_drain();
2587         tlb_gather_mmu(&tlb, mm, start, end);
2588         update_hiwater_rss(mm);
2589         unmap_vmas(&tlb, vma, start, end);
2590         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2591                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2592         tlb_finish_mmu(&tlb, start, end);
2593 }
2594
2595 /*
2596  * Create a list of vma's touched by the unmap, removing them from the mm's
2597  * vma list as we go..
2598  */
2599 static void
2600 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2601         struct vm_area_struct *prev, unsigned long end)
2602 {
2603         struct vm_area_struct **insertion_point;
2604         struct vm_area_struct *tail_vma = NULL;
2605
2606         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2607         vma->vm_prev = NULL;
2608         do {
2609                 vma_rb_erase(vma, &mm->mm_rb);
2610                 mm->map_count--;
2611                 tail_vma = vma;
2612                 vma = vma->vm_next;
2613         } while (vma && vma->vm_start < end);
2614         *insertion_point = vma;
2615         if (vma) {
2616                 vma->vm_prev = prev;
2617                 vma_gap_update(vma);
2618         } else
2619                 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2620         tail_vma->vm_next = NULL;
2621
2622         /* Kill the cache */
2623         vmacache_invalidate(mm);
2624 }
2625
2626 /*
2627  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2628  * has already been checked or doesn't make sense to fail.
2629  */
2630 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2631                 unsigned long addr, int new_below)
2632 {
2633         struct vm_area_struct *new;
2634         int err;
2635
2636         if (vma->vm_ops && vma->vm_ops->split) {
2637                 err = vma->vm_ops->split(vma, addr);
2638                 if (err)
2639                         return err;
2640         }
2641
2642         new = vm_area_dup(vma);
2643         if (!new)
2644                 return -ENOMEM;
2645
2646         if (new_below)
2647                 new->vm_end = addr;
2648         else {
2649                 new->vm_start = addr;
2650                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2651         }
2652
2653         err = vma_dup_policy(vma, new);
2654         if (err)
2655                 goto out_free_vma;
2656
2657         err = anon_vma_clone(new, vma);
2658         if (err)
2659                 goto out_free_mpol;
2660
2661         if (new->vm_file)
2662                 get_file(new->vm_file);
2663
2664         if (new->vm_ops && new->vm_ops->open)
2665                 new->vm_ops->open(new);
2666
2667         if (new_below)
2668                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2669                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2670         else
2671                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2672
2673         /* Success. */
2674         if (!err)
2675                 return 0;
2676
2677         /* Clean everything up if vma_adjust failed. */
2678         if (new->vm_ops && new->vm_ops->close)
2679                 new->vm_ops->close(new);
2680         if (new->vm_file)
2681                 fput(new->vm_file);
2682         unlink_anon_vmas(new);
2683  out_free_mpol:
2684         mpol_put(vma_policy(new));
2685  out_free_vma:
2686         vm_area_free(new);
2687         return err;
2688 }
2689
2690 /*
2691  * Split a vma into two pieces at address 'addr', a new vma is allocated
2692  * either for the first part or the tail.
2693  */
2694 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2695               unsigned long addr, int new_below)
2696 {
2697         if (mm->map_count >= sysctl_max_map_count)
2698                 return -ENOMEM;
2699
2700         return __split_vma(mm, vma, addr, new_below);
2701 }
2702
2703 /* Munmap is split into 2 main parts -- this part which finds
2704  * what needs doing, and the areas themselves, which do the
2705  * work.  This now handles partial unmappings.
2706  * Jeremy Fitzhardinge <jeremy@goop.org>
2707  */
2708 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2709                 struct list_head *uf, bool downgrade)
2710 {
2711         unsigned long end;
2712         struct vm_area_struct *vma, *prev, *last;
2713
2714         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2715                 return -EINVAL;
2716
2717         len = PAGE_ALIGN(len);
2718         end = start + len;
2719         if (len == 0)
2720                 return -EINVAL;
2721
2722         /*
2723          * arch_unmap() might do unmaps itself.  It must be called
2724          * and finish any rbtree manipulation before this code
2725          * runs and also starts to manipulate the rbtree.
2726          */
2727         arch_unmap(mm, start, end);
2728
2729         /* Find the first overlapping VMA */
2730         vma = find_vma(mm, start);
2731         if (!vma)
2732                 return 0;
2733         prev = vma->vm_prev;
2734         /* we have  start < vma->vm_end  */
2735
2736         /* if it doesn't overlap, we have nothing.. */
2737         if (vma->vm_start >= end)
2738                 return 0;
2739
2740         /*
2741          * If we need to split any vma, do it now to save pain later.
2742          *
2743          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2744          * unmapped vm_area_struct will remain in use: so lower split_vma
2745          * places tmp vma above, and higher split_vma places tmp vma below.
2746          */
2747         if (start > vma->vm_start) {
2748                 int error;
2749
2750                 /*
2751                  * Make sure that map_count on return from munmap() will
2752                  * not exceed its limit; but let map_count go just above
2753                  * its limit temporarily, to help free resources as expected.
2754                  */
2755                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2756                         return -ENOMEM;
2757
2758                 error = __split_vma(mm, vma, start, 0);
2759                 if (error)
2760                         return error;
2761                 prev = vma;
2762         }
2763
2764         /* Does it split the last one? */
2765         last = find_vma(mm, end);
2766         if (last && end > last->vm_start) {
2767                 int error = __split_vma(mm, last, end, 1);
2768                 if (error)
2769                         return error;
2770         }
2771         vma = prev ? prev->vm_next : mm->mmap;
2772
2773         if (unlikely(uf)) {
2774                 /*
2775                  * If userfaultfd_unmap_prep returns an error the vmas
2776                  * will remain splitted, but userland will get a
2777                  * highly unexpected error anyway. This is no
2778                  * different than the case where the first of the two
2779                  * __split_vma fails, but we don't undo the first
2780                  * split, despite we could. This is unlikely enough
2781                  * failure that it's not worth optimizing it for.
2782                  */
2783                 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2784                 if (error)
2785                         return error;
2786         }
2787
2788         /*
2789          * unlock any mlock()ed ranges before detaching vmas
2790          */
2791         if (mm->locked_vm) {
2792                 struct vm_area_struct *tmp = vma;
2793                 while (tmp && tmp->vm_start < end) {
2794                         if (tmp->vm_flags & VM_LOCKED) {
2795                                 mm->locked_vm -= vma_pages(tmp);
2796                                 munlock_vma_pages_all(tmp);
2797                         }
2798
2799                         tmp = tmp->vm_next;
2800                 }
2801         }
2802
2803         /* Detach vmas from rbtree */
2804         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2805
2806         if (downgrade)
2807                 downgrade_write(&mm->mmap_sem);
2808
2809         unmap_region(mm, vma, prev, start, end);
2810
2811         /* Fix up all other VM information */
2812         remove_vma_list(mm, vma);
2813
2814         return downgrade ? 1 : 0;
2815 }
2816
2817 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2818               struct list_head *uf)
2819 {
2820         return __do_munmap(mm, start, len, uf, false);
2821 }
2822
2823 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2824 {
2825         int ret;
2826         struct mm_struct *mm = current->mm;
2827         LIST_HEAD(uf);
2828
2829         if (down_write_killable(&mm->mmap_sem))
2830                 return -EINTR;
2831
2832         ret = __do_munmap(mm, start, len, &uf, downgrade);
2833         /*
2834          * Returning 1 indicates mmap_sem is downgraded.
2835          * But 1 is not legal return value of vm_munmap() and munmap(), reset
2836          * it to 0 before return.
2837          */
2838         if (ret == 1) {
2839                 up_read(&mm->mmap_sem);
2840                 ret = 0;
2841         } else
2842                 up_write(&mm->mmap_sem);
2843
2844         userfaultfd_unmap_complete(mm, &uf);
2845         return ret;
2846 }
2847
2848 int vm_munmap(unsigned long start, size_t len)
2849 {
2850         return __vm_munmap(start, len, false);
2851 }
2852 EXPORT_SYMBOL(vm_munmap);
2853
2854 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2855 {
2856         addr = untagged_addr(addr);
2857         profile_munmap(addr);
2858         return __vm_munmap(addr, len, true);
2859 }
2860
2861
2862 /*
2863  * Emulation of deprecated remap_file_pages() syscall.
2864  */
2865 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2866                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2867 {
2868
2869         struct mm_struct *mm = current->mm;
2870         struct vm_area_struct *vma;
2871         unsigned long populate = 0;
2872         unsigned long ret = -EINVAL;
2873         struct file *file;
2874
2875         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2876                      current->comm, current->pid);
2877
2878         if (prot)
2879                 return ret;
2880         start = start & PAGE_MASK;
2881         size = size & PAGE_MASK;
2882
2883         if (start + size <= start)
2884                 return ret;
2885
2886         /* Does pgoff wrap? */
2887         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2888                 return ret;
2889
2890         if (down_write_killable(&mm->mmap_sem))
2891                 return -EINTR;
2892
2893         vma = find_vma(mm, start);
2894
2895         if (!vma || !(vma->vm_flags & VM_SHARED))
2896                 goto out;
2897
2898         if (start < vma->vm_start)
2899                 goto out;
2900
2901         if (start + size > vma->vm_end) {
2902                 struct vm_area_struct *next;
2903
2904                 for (next = vma->vm_next; next; next = next->vm_next) {
2905                         /* hole between vmas ? */
2906                         if (next->vm_start != next->vm_prev->vm_end)
2907                                 goto out;
2908
2909                         if (next->vm_file != vma->vm_file)
2910                                 goto out;
2911
2912                         if (next->vm_flags != vma->vm_flags)
2913                                 goto out;
2914
2915                         if (start + size <= next->vm_end)
2916                                 break;
2917                 }
2918
2919                 if (!next)
2920                         goto out;
2921         }
2922
2923         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2924         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2925         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2926
2927         flags &= MAP_NONBLOCK;
2928         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2929         if (vma->vm_flags & VM_LOCKED) {
2930                 struct vm_area_struct *tmp;
2931                 flags |= MAP_LOCKED;
2932
2933                 /* drop PG_Mlocked flag for over-mapped range */
2934                 for (tmp = vma; tmp->vm_start >= start + size;
2935                                 tmp = tmp->vm_next) {
2936                         /*
2937                          * Split pmd and munlock page on the border
2938                          * of the range.
2939                          */
2940                         vma_adjust_trans_huge(tmp, start, start + size, 0);
2941
2942                         munlock_vma_pages_range(tmp,
2943                                         max(tmp->vm_start, start),
2944                                         min(tmp->vm_end, start + size));
2945                 }
2946         }
2947
2948         file = get_file(vma->vm_file);
2949         ret = do_mmap_pgoff(vma->vm_file, start, size,
2950                         prot, flags, pgoff, &populate, NULL);
2951         fput(file);
2952 out:
2953         up_write(&mm->mmap_sem);
2954         if (populate)
2955                 mm_populate(ret, populate);
2956         if (!IS_ERR_VALUE(ret))
2957                 ret = 0;
2958         return ret;
2959 }
2960
2961 /*
2962  *  this is really a simplified "do_mmap".  it only handles
2963  *  anonymous maps.  eventually we may be able to do some
2964  *  brk-specific accounting here.
2965  */
2966 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
2967 {
2968         struct mm_struct *mm = current->mm;
2969         struct vm_area_struct *vma, *prev;
2970         struct rb_node **rb_link, *rb_parent;
2971         pgoff_t pgoff = addr >> PAGE_SHIFT;
2972         int error;
2973         unsigned long mapped_addr;
2974
2975         /* Until we need other flags, refuse anything except VM_EXEC. */
2976         if ((flags & (~VM_EXEC)) != 0)
2977                 return -EINVAL;
2978         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2979
2980         mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2981         if (IS_ERR_VALUE(mapped_addr))
2982                 return mapped_addr;
2983
2984         error = mlock_future_check(mm, mm->def_flags, len);
2985         if (error)
2986                 return error;
2987
2988         /*
2989          * Clear old maps.  this also does some error checking for us
2990          */
2991         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2992                               &rb_parent)) {
2993                 if (do_munmap(mm, addr, len, uf))
2994                         return -ENOMEM;
2995         }
2996
2997         /* Check against address space limits *after* clearing old maps... */
2998         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2999                 return -ENOMEM;
3000
3001         if (mm->map_count > sysctl_max_map_count)
3002                 return -ENOMEM;
3003
3004         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3005                 return -ENOMEM;
3006
3007         /* Can we just expand an old private anonymous mapping? */
3008         vma = vma_merge(mm, prev, addr, addr + len, flags,
3009                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3010         if (vma)
3011                 goto out;
3012
3013         /*
3014          * create a vma struct for an anonymous mapping
3015          */
3016         vma = vm_area_alloc(mm);
3017         if (!vma) {
3018                 vm_unacct_memory(len >> PAGE_SHIFT);
3019                 return -ENOMEM;
3020         }
3021
3022         vma_set_anonymous(vma);
3023         vma->vm_start = addr;
3024         vma->vm_end = addr + len;
3025         vma->vm_pgoff = pgoff;
3026         vma->vm_flags = flags;
3027         vma->vm_page_prot = vm_get_page_prot(flags);
3028         vma_link(mm, vma, prev, rb_link, rb_parent);
3029 out:
3030         perf_event_mmap(vma);
3031         mm->total_vm += len >> PAGE_SHIFT;
3032         mm->data_vm += len >> PAGE_SHIFT;
3033         if (flags & VM_LOCKED)
3034                 mm->locked_vm += (len >> PAGE_SHIFT);
3035         vma->vm_flags |= VM_SOFTDIRTY;
3036         return 0;
3037 }
3038
3039 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3040 {
3041         struct mm_struct *mm = current->mm;
3042         unsigned long len;
3043         int ret;
3044         bool populate;
3045         LIST_HEAD(uf);
3046
3047         len = PAGE_ALIGN(request);
3048         if (len < request)
3049                 return -ENOMEM;
3050         if (!len)
3051                 return 0;
3052
3053         if (down_write_killable(&mm->mmap_sem))
3054                 return -EINTR;
3055
3056         ret = do_brk_flags(addr, len, flags, &uf);
3057         populate = ((mm->def_flags & VM_LOCKED) != 0);
3058         up_write(&mm->mmap_sem);
3059         userfaultfd_unmap_complete(mm, &uf);
3060         if (populate && !ret)
3061                 mm_populate(addr, len);
3062         return ret;
3063 }
3064 EXPORT_SYMBOL(vm_brk_flags);
3065
3066 int vm_brk(unsigned long addr, unsigned long len)
3067 {
3068         return vm_brk_flags(addr, len, 0);
3069 }
3070 EXPORT_SYMBOL(vm_brk);
3071
3072 /* Release all mmaps. */
3073 void exit_mmap(struct mm_struct *mm)
3074 {
3075         struct mmu_gather tlb;
3076         struct vm_area_struct *vma;
3077         unsigned long nr_accounted = 0;
3078
3079         /* mm's last user has gone, and its about to be pulled down */
3080         mmu_notifier_release(mm);
3081
3082         if (unlikely(mm_is_oom_victim(mm))) {
3083                 /*
3084                  * Manually reap the mm to free as much memory as possible.
3085                  * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3086                  * this mm from further consideration.  Taking mm->mmap_sem for
3087                  * write after setting MMF_OOM_SKIP will guarantee that the oom
3088                  * reaper will not run on this mm again after mmap_sem is
3089                  * dropped.
3090                  *
3091                  * Nothing can be holding mm->mmap_sem here and the above call
3092                  * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3093                  * __oom_reap_task_mm() will not block.
3094                  *
3095                  * This needs to be done before calling munlock_vma_pages_all(),
3096                  * which clears VM_LOCKED, otherwise the oom reaper cannot
3097                  * reliably test it.
3098                  */
3099                 (void)__oom_reap_task_mm(mm);
3100
3101                 set_bit(MMF_OOM_SKIP, &mm->flags);
3102                 down_write(&mm->mmap_sem);
3103                 up_write(&mm->mmap_sem);
3104         }
3105
3106         if (mm->locked_vm) {
3107                 vma = mm->mmap;
3108                 while (vma) {
3109                         if (vma->vm_flags & VM_LOCKED)
3110                                 munlock_vma_pages_all(vma);
3111                         vma = vma->vm_next;
3112                 }
3113         }
3114
3115         arch_exit_mmap(mm);
3116
3117         vma = mm->mmap;
3118         if (!vma)       /* Can happen if dup_mmap() received an OOM */
3119                 return;
3120
3121         lru_add_drain();
3122         flush_cache_mm(mm);
3123         tlb_gather_mmu(&tlb, mm, 0, -1);
3124         /* update_hiwater_rss(mm) here? but nobody should be looking */
3125         /* Use -1 here to ensure all VMAs in the mm are unmapped */
3126         unmap_vmas(&tlb, vma, 0, -1);
3127         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3128         tlb_finish_mmu(&tlb, 0, -1);
3129
3130         /*
3131          * Walk the list again, actually closing and freeing it,
3132          * with preemption enabled, without holding any MM locks.
3133          */
3134         while (vma) {
3135                 if (vma->vm_flags & VM_ACCOUNT)
3136                         nr_accounted += vma_pages(vma);
3137                 vma = remove_vma(vma);
3138         }
3139         vm_unacct_memory(nr_accounted);
3140 }
3141
3142 /* Insert vm structure into process list sorted by address
3143  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3144  * then i_mmap_rwsem is taken here.
3145  */
3146 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3147 {
3148         struct vm_area_struct *prev;
3149         struct rb_node **rb_link, *rb_parent;
3150
3151         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3152                            &prev, &rb_link, &rb_parent))
3153                 return -ENOMEM;
3154         if ((vma->vm_flags & VM_ACCOUNT) &&
3155              security_vm_enough_memory_mm(mm, vma_pages(vma)))
3156                 return -ENOMEM;
3157
3158         /*
3159          * The vm_pgoff of a purely anonymous vma should be irrelevant
3160          * until its first write fault, when page's anon_vma and index
3161          * are set.  But now set the vm_pgoff it will almost certainly
3162          * end up with (unless mremap moves it elsewhere before that
3163          * first wfault), so /proc/pid/maps tells a consistent story.
3164          *
3165          * By setting it to reflect the virtual start address of the
3166          * vma, merges and splits can happen in a seamless way, just
3167          * using the existing file pgoff checks and manipulations.
3168          * Similarly in do_mmap_pgoff and in do_brk.
3169          */
3170         if (vma_is_anonymous(vma)) {
3171                 BUG_ON(vma->anon_vma);
3172                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3173         }
3174
3175         vma_link(mm, vma, prev, rb_link, rb_parent);
3176         return 0;
3177 }
3178
3179 /*
3180  * Copy the vma structure to a new location in the same mm,
3181  * prior to moving page table entries, to effect an mremap move.
3182  */
3183 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3184         unsigned long addr, unsigned long len, pgoff_t pgoff,
3185         bool *need_rmap_locks)
3186 {
3187         struct vm_area_struct *vma = *vmap;
3188         unsigned long vma_start = vma->vm_start;
3189         struct mm_struct *mm = vma->vm_mm;
3190         struct vm_area_struct *new_vma, *prev;
3191         struct rb_node **rb_link, *rb_parent;
3192         bool faulted_in_anon_vma = true;
3193
3194         /*
3195          * If anonymous vma has not yet been faulted, update new pgoff
3196          * to match new location, to increase its chance of merging.
3197          */
3198         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3199                 pgoff = addr >> PAGE_SHIFT;
3200                 faulted_in_anon_vma = false;
3201         }
3202
3203         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3204                 return NULL;    /* should never get here */
3205         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3206                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3207                             vma->vm_userfaultfd_ctx);
3208         if (new_vma) {
3209                 /*
3210                  * Source vma may have been merged into new_vma
3211                  */
3212                 if (unlikely(vma_start >= new_vma->vm_start &&
3213                              vma_start < new_vma->vm_end)) {
3214                         /*
3215                          * The only way we can get a vma_merge with
3216                          * self during an mremap is if the vma hasn't
3217                          * been faulted in yet and we were allowed to
3218                          * reset the dst vma->vm_pgoff to the
3219                          * destination address of the mremap to allow
3220                          * the merge to happen. mremap must change the
3221                          * vm_pgoff linearity between src and dst vmas
3222                          * (in turn preventing a vma_merge) to be
3223                          * safe. It is only safe to keep the vm_pgoff
3224                          * linear if there are no pages mapped yet.
3225                          */
3226                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3227                         *vmap = vma = new_vma;
3228                 }
3229                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3230         } else {
3231                 new_vma = vm_area_dup(vma);
3232                 if (!new_vma)
3233                         goto out;
3234                 new_vma->vm_start = addr;
3235                 new_vma->vm_end = addr + len;
3236                 new_vma->vm_pgoff = pgoff;
3237                 if (vma_dup_policy(vma, new_vma))
3238                         goto out_free_vma;
3239                 if (anon_vma_clone(new_vma, vma))
3240                         goto out_free_mempol;
3241                 if (new_vma->vm_file)
3242                         get_file(new_vma->vm_file);
3243                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3244                         new_vma->vm_ops->open(new_vma);
3245                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3246                 *need_rmap_locks = false;
3247         }
3248         return new_vma;
3249
3250 out_free_mempol:
3251         mpol_put(vma_policy(new_vma));
3252 out_free_vma:
3253         vm_area_free(new_vma);
3254 out:
3255         return NULL;
3256 }
3257
3258 /*
3259  * Return true if the calling process may expand its vm space by the passed
3260  * number of pages
3261  */
3262 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3263 {
3264         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3265                 return false;
3266
3267         if (is_data_mapping(flags) &&
3268             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3269                 /* Workaround for Valgrind */
3270                 if (rlimit(RLIMIT_DATA) == 0 &&
3271                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3272                         return true;
3273
3274                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3275                              current->comm, current->pid,
3276                              (mm->data_vm + npages) << PAGE_SHIFT,
3277                              rlimit(RLIMIT_DATA),
3278                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3279
3280                 if (!ignore_rlimit_data)
3281                         return false;
3282         }
3283
3284         return true;
3285 }
3286
3287 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3288 {
3289         mm->total_vm += npages;
3290
3291         if (is_exec_mapping(flags))
3292                 mm->exec_vm += npages;
3293         else if (is_stack_mapping(flags))
3294                 mm->stack_vm += npages;
3295         else if (is_data_mapping(flags))
3296                 mm->data_vm += npages;
3297 }
3298
3299 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3300
3301 /*
3302  * Having a close hook prevents vma merging regardless of flags.
3303  */
3304 static void special_mapping_close(struct vm_area_struct *vma)
3305 {
3306 }
3307
3308 static const char *special_mapping_name(struct vm_area_struct *vma)
3309 {
3310         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3311 }
3312
3313 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3314 {
3315         struct vm_special_mapping *sm = new_vma->vm_private_data;
3316
3317         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3318                 return -EFAULT;
3319
3320         if (sm->mremap)
3321                 return sm->mremap(sm, new_vma);
3322
3323         return 0;
3324 }
3325
3326 static const struct vm_operations_struct special_mapping_vmops = {
3327         .close = special_mapping_close,
3328         .fault = special_mapping_fault,
3329         .mremap = special_mapping_mremap,
3330         .name = special_mapping_name,
3331         /* vDSO code relies that VVAR can't be accessed remotely */
3332         .access = NULL,
3333 };
3334
3335 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3336         .close = special_mapping_close,
3337         .fault = special_mapping_fault,
3338 };
3339
3340 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3341 {
3342         struct vm_area_struct *vma = vmf->vma;
3343         pgoff_t pgoff;
3344         struct page **pages;
3345
3346         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3347                 pages = vma->vm_private_data;
3348         } else {
3349                 struct vm_special_mapping *sm = vma->vm_private_data;
3350
3351                 if (sm->fault)
3352                         return sm->fault(sm, vmf->vma, vmf);
3353
3354                 pages = sm->pages;
3355         }
3356
3357         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3358                 pgoff--;
3359
3360         if (*pages) {
3361                 struct page *page = *pages;
3362                 get_page(page);
3363                 vmf->page = page;
3364                 return 0;
3365         }
3366
3367         return VM_FAULT_SIGBUS;
3368 }
3369
3370 static struct vm_area_struct *__install_special_mapping(
3371         struct mm_struct *mm,
3372         unsigned long addr, unsigned long len,
3373         unsigned long vm_flags, void *priv,
3374         const struct vm_operations_struct *ops)
3375 {
3376         int ret;
3377         struct vm_area_struct *vma;
3378
3379         vma = vm_area_alloc(mm);
3380         if (unlikely(vma == NULL))
3381                 return ERR_PTR(-ENOMEM);
3382
3383         vma->vm_start = addr;
3384         vma->vm_end = addr + len;
3385
3386         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
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_sem 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_sem);
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_sem);
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_sem in write mode before calling
3493  * mm_take_all_locks(). The caller isn't allowed to release the
3494  * mmap_sem until mm_drop_all_locks() returns.
3495  *
3496  * mmap_sem 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         BUG_ON(down_read_trylock(&mm->mmap_sem));
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_sem 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         BUG_ON(down_read_trylock(&mm->mmap_sem));
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);