Merge tag 'reset-fixes-for-4.18' of git://git.pengutronix.de/git/pza/linux into next...
[platform/kernel/linux-rpi.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
14 #include <linux/mm.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
24 #include <linux/fs.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/shmem_fs.h>
29 #include <linux/profile.h>
30 #include <linux/export.h>
31 #include <linux/mount.h>
32 #include <linux/mempolicy.h>
33 #include <linux/rmap.h>
34 #include <linux/mmu_notifier.h>
35 #include <linux/mmdebug.h>
36 #include <linux/perf_event.h>
37 #include <linux/audit.h>
38 #include <linux/khugepaged.h>
39 #include <linux/uprobes.h>
40 #include <linux/rbtree_augmented.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/moduleparam.h>
46 #include <linux/pkeys.h>
47 #include <linux/oom.h>
48
49 #include <linux/uaccess.h>
50 #include <asm/cacheflush.h>
51 #include <asm/tlb.h>
52 #include <asm/mmu_context.h>
53
54 #include "internal.h"
55
56 #ifndef arch_mmap_check
57 #define arch_mmap_check(addr, len, flags)       (0)
58 #endif
59
60 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
61 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
62 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
63 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
64 #endif
65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
66 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
67 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
68 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
69 #endif
70
71 static bool ignore_rlimit_data;
72 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
73
74 static void unmap_region(struct mm_struct *mm,
75                 struct vm_area_struct *vma, struct vm_area_struct *prev,
76                 unsigned long start, unsigned long end);
77
78 /* description of effects of mapping type and prot in current implementation.
79  * this is due to the limited x86 page protection hardware.  The expected
80  * behavior is in parens:
81  *
82  * map_type     prot
83  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
84  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
85  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
86  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
87  *
88  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
89  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
90  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
91  *
92  * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
93  * MAP_PRIVATE:
94  *                                                              r: (no) no
95  *                                                              w: (no) no
96  *                                                              x: (yes) yes
97  */
98 pgprot_t protection_map[16] __ro_after_init = {
99         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
100         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
101 };
102
103 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
104 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
105 {
106         return prot;
107 }
108 #endif
109
110 pgprot_t vm_get_page_prot(unsigned long vm_flags)
111 {
112         pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
113                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
114                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
115
116         return arch_filter_pgprot(ret);
117 }
118 EXPORT_SYMBOL(vm_get_page_prot);
119
120 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
121 {
122         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
123 }
124
125 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
126 void vma_set_page_prot(struct vm_area_struct *vma)
127 {
128         unsigned long vm_flags = vma->vm_flags;
129         pgprot_t vm_page_prot;
130
131         vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
132         if (vma_wants_writenotify(vma, vm_page_prot)) {
133                 vm_flags &= ~VM_SHARED;
134                 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
135         }
136         /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
137         WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
138 }
139
140 /*
141  * Requires inode->i_mapping->i_mmap_rwsem
142  */
143 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
144                 struct file *file, struct address_space *mapping)
145 {
146         if (vma->vm_flags & VM_DENYWRITE)
147                 atomic_inc(&file_inode(file)->i_writecount);
148         if (vma->vm_flags & VM_SHARED)
149                 mapping_unmap_writable(mapping);
150
151         flush_dcache_mmap_lock(mapping);
152         vma_interval_tree_remove(vma, &mapping->i_mmap);
153         flush_dcache_mmap_unlock(mapping);
154 }
155
156 /*
157  * Unlink a file-based vm structure from its interval tree, to hide
158  * vma from rmap and vmtruncate before freeing its page tables.
159  */
160 void unlink_file_vma(struct vm_area_struct *vma)
161 {
162         struct file *file = vma->vm_file;
163
164         if (file) {
165                 struct address_space *mapping = file->f_mapping;
166                 i_mmap_lock_write(mapping);
167                 __remove_shared_vm_struct(vma, file, mapping);
168                 i_mmap_unlock_write(mapping);
169         }
170 }
171
172 /*
173  * Close a vm structure and free it, returning the next.
174  */
175 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
176 {
177         struct vm_area_struct *next = vma->vm_next;
178
179         might_sleep();
180         if (vma->vm_ops && vma->vm_ops->close)
181                 vma->vm_ops->close(vma);
182         if (vma->vm_file)
183                 fput(vma->vm_file);
184         mpol_put(vma_policy(vma));
185         kmem_cache_free(vm_area_cachep, vma);
186         return next;
187 }
188
189 static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf);
190
191 SYSCALL_DEFINE1(brk, unsigned long, brk)
192 {
193         unsigned long retval;
194         unsigned long newbrk, oldbrk;
195         struct mm_struct *mm = current->mm;
196         struct vm_area_struct *next;
197         unsigned long min_brk;
198         bool populate;
199         LIST_HEAD(uf);
200
201         if (down_write_killable(&mm->mmap_sem))
202                 return -EINTR;
203
204 #ifdef CONFIG_COMPAT_BRK
205         /*
206          * CONFIG_COMPAT_BRK can still be overridden by setting
207          * randomize_va_space to 2, which will still cause mm->start_brk
208          * to be arbitrarily shifted
209          */
210         if (current->brk_randomized)
211                 min_brk = mm->start_brk;
212         else
213                 min_brk = mm->end_data;
214 #else
215         min_brk = mm->start_brk;
216 #endif
217         if (brk < min_brk)
218                 goto out;
219
220         /*
221          * Check against rlimit here. If this check is done later after the test
222          * of oldbrk with newbrk then it can escape the test and let the data
223          * segment grow beyond its set limit the in case where the limit is
224          * not page aligned -Ram Gupta
225          */
226         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
227                               mm->end_data, mm->start_data))
228                 goto out;
229
230         newbrk = PAGE_ALIGN(brk);
231         oldbrk = PAGE_ALIGN(mm->brk);
232         if (oldbrk == newbrk)
233                 goto set_brk;
234
235         /* Always allow shrinking brk. */
236         if (brk <= mm->brk) {
237                 if (!do_munmap(mm, newbrk, oldbrk-newbrk, &uf))
238                         goto set_brk;
239                 goto out;
240         }
241
242         /* Check against existing mmap mappings. */
243         next = find_vma(mm, oldbrk);
244         if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
245                 goto out;
246
247         /* Ok, looks good - let it rip. */
248         if (do_brk(oldbrk, newbrk-oldbrk, &uf) < 0)
249                 goto out;
250
251 set_brk:
252         mm->brk = brk;
253         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
254         up_write(&mm->mmap_sem);
255         userfaultfd_unmap_complete(mm, &uf);
256         if (populate)
257                 mm_populate(oldbrk, newbrk - oldbrk);
258         return brk;
259
260 out:
261         retval = mm->brk;
262         up_write(&mm->mmap_sem);
263         return retval;
264 }
265
266 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
267 {
268         unsigned long max, prev_end, subtree_gap;
269
270         /*
271          * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
272          * allow two stack_guard_gaps between them here, and when choosing
273          * an unmapped area; whereas when expanding we only require one.
274          * That's a little inconsistent, but keeps the code here simpler.
275          */
276         max = vm_start_gap(vma);
277         if (vma->vm_prev) {
278                 prev_end = vm_end_gap(vma->vm_prev);
279                 if (max > prev_end)
280                         max -= prev_end;
281                 else
282                         max = 0;
283         }
284         if (vma->vm_rb.rb_left) {
285                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
286                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
287                 if (subtree_gap > max)
288                         max = subtree_gap;
289         }
290         if (vma->vm_rb.rb_right) {
291                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
292                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
293                 if (subtree_gap > max)
294                         max = subtree_gap;
295         }
296         return max;
297 }
298
299 #ifdef CONFIG_DEBUG_VM_RB
300 static int browse_rb(struct mm_struct *mm)
301 {
302         struct rb_root *root = &mm->mm_rb;
303         int i = 0, j, bug = 0;
304         struct rb_node *nd, *pn = NULL;
305         unsigned long prev = 0, pend = 0;
306
307         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308                 struct vm_area_struct *vma;
309                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310                 if (vma->vm_start < prev) {
311                         pr_emerg("vm_start %lx < prev %lx\n",
312                                   vma->vm_start, prev);
313                         bug = 1;
314                 }
315                 if (vma->vm_start < pend) {
316                         pr_emerg("vm_start %lx < pend %lx\n",
317                                   vma->vm_start, pend);
318                         bug = 1;
319                 }
320                 if (vma->vm_start > vma->vm_end) {
321                         pr_emerg("vm_start %lx > vm_end %lx\n",
322                                   vma->vm_start, vma->vm_end);
323                         bug = 1;
324                 }
325                 spin_lock(&mm->page_table_lock);
326                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
327                         pr_emerg("free gap %lx, correct %lx\n",
328                                vma->rb_subtree_gap,
329                                vma_compute_subtree_gap(vma));
330                         bug = 1;
331                 }
332                 spin_unlock(&mm->page_table_lock);
333                 i++;
334                 pn = nd;
335                 prev = vma->vm_start;
336                 pend = vma->vm_end;
337         }
338         j = 0;
339         for (nd = pn; nd; nd = rb_prev(nd))
340                 j++;
341         if (i != j) {
342                 pr_emerg("backwards %d, forwards %d\n", j, i);
343                 bug = 1;
344         }
345         return bug ? -1 : i;
346 }
347
348 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
349 {
350         struct rb_node *nd;
351
352         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
353                 struct vm_area_struct *vma;
354                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
355                 VM_BUG_ON_VMA(vma != ignore &&
356                         vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
357                         vma);
358         }
359 }
360
361 static void validate_mm(struct mm_struct *mm)
362 {
363         int bug = 0;
364         int i = 0;
365         unsigned long highest_address = 0;
366         struct vm_area_struct *vma = mm->mmap;
367
368         while (vma) {
369                 struct anon_vma *anon_vma = vma->anon_vma;
370                 struct anon_vma_chain *avc;
371
372                 if (anon_vma) {
373                         anon_vma_lock_read(anon_vma);
374                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
375                                 anon_vma_interval_tree_verify(avc);
376                         anon_vma_unlock_read(anon_vma);
377                 }
378
379                 highest_address = vm_end_gap(vma);
380                 vma = vma->vm_next;
381                 i++;
382         }
383         if (i != mm->map_count) {
384                 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
385                 bug = 1;
386         }
387         if (highest_address != mm->highest_vm_end) {
388                 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
389                           mm->highest_vm_end, highest_address);
390                 bug = 1;
391         }
392         i = browse_rb(mm);
393         if (i != mm->map_count) {
394                 if (i != -1)
395                         pr_emerg("map_count %d rb %d\n", mm->map_count, i);
396                 bug = 1;
397         }
398         VM_BUG_ON_MM(bug, mm);
399 }
400 #else
401 #define validate_mm_rb(root, ignore) do { } while (0)
402 #define validate_mm(mm) do { } while (0)
403 #endif
404
405 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
406                      unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
407
408 /*
409  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
410  * vma->vm_prev->vm_end values changed, without modifying the vma's position
411  * in the rbtree.
412  */
413 static void vma_gap_update(struct vm_area_struct *vma)
414 {
415         /*
416          * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
417          * function that does exacltly what we want.
418          */
419         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
420 }
421
422 static inline void vma_rb_insert(struct vm_area_struct *vma,
423                                  struct rb_root *root)
424 {
425         /* All rb_subtree_gap values must be consistent prior to insertion */
426         validate_mm_rb(root, NULL);
427
428         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
429 }
430
431 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
432 {
433         /*
434          * Note rb_erase_augmented is a fairly large inline function,
435          * so make sure we instantiate it only once with our desired
436          * augmented rbtree callbacks.
437          */
438         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
439 }
440
441 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
442                                                 struct rb_root *root,
443                                                 struct vm_area_struct *ignore)
444 {
445         /*
446          * All rb_subtree_gap values must be consistent prior to erase,
447          * with the possible exception of the "next" vma being erased if
448          * next->vm_start was reduced.
449          */
450         validate_mm_rb(root, ignore);
451
452         __vma_rb_erase(vma, root);
453 }
454
455 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
456                                          struct rb_root *root)
457 {
458         /*
459          * All rb_subtree_gap values must be consistent prior to erase,
460          * with the possible exception of the vma being erased.
461          */
462         validate_mm_rb(root, vma);
463
464         __vma_rb_erase(vma, root);
465 }
466
467 /*
468  * vma has some anon_vma assigned, and is already inserted on that
469  * anon_vma's interval trees.
470  *
471  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
472  * vma must be removed from the anon_vma's interval trees using
473  * anon_vma_interval_tree_pre_update_vma().
474  *
475  * After the update, the vma will be reinserted using
476  * anon_vma_interval_tree_post_update_vma().
477  *
478  * The entire update must be protected by exclusive mmap_sem and by
479  * the root anon_vma's mutex.
480  */
481 static inline void
482 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
483 {
484         struct anon_vma_chain *avc;
485
486         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
487                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
488 }
489
490 static inline void
491 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
492 {
493         struct anon_vma_chain *avc;
494
495         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
496                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
497 }
498
499 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
500                 unsigned long end, struct vm_area_struct **pprev,
501                 struct rb_node ***rb_link, struct rb_node **rb_parent)
502 {
503         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
504
505         __rb_link = &mm->mm_rb.rb_node;
506         rb_prev = __rb_parent = NULL;
507
508         while (*__rb_link) {
509                 struct vm_area_struct *vma_tmp;
510
511                 __rb_parent = *__rb_link;
512                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
513
514                 if (vma_tmp->vm_end > addr) {
515                         /* Fail if an existing vma overlaps the area */
516                         if (vma_tmp->vm_start < end)
517                                 return -ENOMEM;
518                         __rb_link = &__rb_parent->rb_left;
519                 } else {
520                         rb_prev = __rb_parent;
521                         __rb_link = &__rb_parent->rb_right;
522                 }
523         }
524
525         *pprev = NULL;
526         if (rb_prev)
527                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
528         *rb_link = __rb_link;
529         *rb_parent = __rb_parent;
530         return 0;
531 }
532
533 static unsigned long count_vma_pages_range(struct mm_struct *mm,
534                 unsigned long addr, unsigned long end)
535 {
536         unsigned long nr_pages = 0;
537         struct vm_area_struct *vma;
538
539         /* Find first overlaping mapping */
540         vma = find_vma_intersection(mm, addr, end);
541         if (!vma)
542                 return 0;
543
544         nr_pages = (min(end, vma->vm_end) -
545                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
546
547         /* Iterate over the rest of the overlaps */
548         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
549                 unsigned long overlap_len;
550
551                 if (vma->vm_start > end)
552                         break;
553
554                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
555                 nr_pages += overlap_len >> PAGE_SHIFT;
556         }
557
558         return nr_pages;
559 }
560
561 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
562                 struct rb_node **rb_link, struct rb_node *rb_parent)
563 {
564         /* Update tracking information for the gap following the new vma. */
565         if (vma->vm_next)
566                 vma_gap_update(vma->vm_next);
567         else
568                 mm->highest_vm_end = vm_end_gap(vma);
569
570         /*
571          * vma->vm_prev wasn't known when we followed the rbtree to find the
572          * correct insertion point for that vma. As a result, we could not
573          * update the vma vm_rb parents rb_subtree_gap values on the way down.
574          * So, we first insert the vma with a zero rb_subtree_gap value
575          * (to be consistent with what we did on the way down), and then
576          * immediately update the gap to the correct value. Finally we
577          * rebalance the rbtree after all augmented values have been set.
578          */
579         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
580         vma->rb_subtree_gap = 0;
581         vma_gap_update(vma);
582         vma_rb_insert(vma, &mm->mm_rb);
583 }
584
585 static void __vma_link_file(struct vm_area_struct *vma)
586 {
587         struct file *file;
588
589         file = vma->vm_file;
590         if (file) {
591                 struct address_space *mapping = file->f_mapping;
592
593                 if (vma->vm_flags & VM_DENYWRITE)
594                         atomic_dec(&file_inode(file)->i_writecount);
595                 if (vma->vm_flags & VM_SHARED)
596                         atomic_inc(&mapping->i_mmap_writable);
597
598                 flush_dcache_mmap_lock(mapping);
599                 vma_interval_tree_insert(vma, &mapping->i_mmap);
600                 flush_dcache_mmap_unlock(mapping);
601         }
602 }
603
604 static void
605 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
606         struct vm_area_struct *prev, struct rb_node **rb_link,
607         struct rb_node *rb_parent)
608 {
609         __vma_link_list(mm, vma, prev, rb_parent);
610         __vma_link_rb(mm, vma, rb_link, rb_parent);
611 }
612
613 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
614                         struct vm_area_struct *prev, struct rb_node **rb_link,
615                         struct rb_node *rb_parent)
616 {
617         struct address_space *mapping = NULL;
618
619         if (vma->vm_file) {
620                 mapping = vma->vm_file->f_mapping;
621                 i_mmap_lock_write(mapping);
622         }
623
624         __vma_link(mm, vma, prev, rb_link, rb_parent);
625         __vma_link_file(vma);
626
627         if (mapping)
628                 i_mmap_unlock_write(mapping);
629
630         mm->map_count++;
631         validate_mm(mm);
632 }
633
634 /*
635  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
636  * mm's list and rbtree.  It has already been inserted into the interval tree.
637  */
638 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
639 {
640         struct vm_area_struct *prev;
641         struct rb_node **rb_link, *rb_parent;
642
643         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
644                            &prev, &rb_link, &rb_parent))
645                 BUG();
646         __vma_link(mm, vma, prev, rb_link, rb_parent);
647         mm->map_count++;
648 }
649
650 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
651                                                 struct vm_area_struct *vma,
652                                                 struct vm_area_struct *prev,
653                                                 bool has_prev,
654                                                 struct vm_area_struct *ignore)
655 {
656         struct vm_area_struct *next;
657
658         vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
659         next = vma->vm_next;
660         if (has_prev)
661                 prev->vm_next = next;
662         else {
663                 prev = vma->vm_prev;
664                 if (prev)
665                         prev->vm_next = next;
666                 else
667                         mm->mmap = next;
668         }
669         if (next)
670                 next->vm_prev = prev;
671
672         /* Kill the cache */
673         vmacache_invalidate(mm);
674 }
675
676 static inline void __vma_unlink_prev(struct mm_struct *mm,
677                                      struct vm_area_struct *vma,
678                                      struct vm_area_struct *prev)
679 {
680         __vma_unlink_common(mm, vma, prev, true, vma);
681 }
682
683 /*
684  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
685  * is already present in an i_mmap tree without adjusting the tree.
686  * The following helper function should be used when such adjustments
687  * are necessary.  The "insert" vma (if any) is to be inserted
688  * before we drop the necessary locks.
689  */
690 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
691         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
692         struct vm_area_struct *expand)
693 {
694         struct mm_struct *mm = vma->vm_mm;
695         struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
696         struct address_space *mapping = NULL;
697         struct rb_root_cached *root = NULL;
698         struct anon_vma *anon_vma = NULL;
699         struct file *file = vma->vm_file;
700         bool start_changed = false, end_changed = false;
701         long adjust_next = 0;
702         int remove_next = 0;
703
704         if (next && !insert) {
705                 struct vm_area_struct *exporter = NULL, *importer = NULL;
706
707                 if (end >= next->vm_end) {
708                         /*
709                          * vma expands, overlapping all the next, and
710                          * perhaps the one after too (mprotect case 6).
711                          * The only other cases that gets here are
712                          * case 1, case 7 and case 8.
713                          */
714                         if (next == expand) {
715                                 /*
716                                  * The only case where we don't expand "vma"
717                                  * and we expand "next" instead is case 8.
718                                  */
719                                 VM_WARN_ON(end != next->vm_end);
720                                 /*
721                                  * remove_next == 3 means we're
722                                  * removing "vma" and that to do so we
723                                  * swapped "vma" and "next".
724                                  */
725                                 remove_next = 3;
726                                 VM_WARN_ON(file != next->vm_file);
727                                 swap(vma, next);
728                         } else {
729                                 VM_WARN_ON(expand != vma);
730                                 /*
731                                  * case 1, 6, 7, remove_next == 2 is case 6,
732                                  * remove_next == 1 is case 1 or 7.
733                                  */
734                                 remove_next = 1 + (end > next->vm_end);
735                                 VM_WARN_ON(remove_next == 2 &&
736                                            end != next->vm_next->vm_end);
737                                 VM_WARN_ON(remove_next == 1 &&
738                                            end != next->vm_end);
739                                 /* trim end to next, for case 6 first pass */
740                                 end = next->vm_end;
741                         }
742
743                         exporter = next;
744                         importer = vma;
745
746                         /*
747                          * If next doesn't have anon_vma, import from vma after
748                          * next, if the vma overlaps with it.
749                          */
750                         if (remove_next == 2 && !next->anon_vma)
751                                 exporter = next->vm_next;
752
753                 } else if (end > next->vm_start) {
754                         /*
755                          * vma expands, overlapping part of the next:
756                          * mprotect case 5 shifting the boundary up.
757                          */
758                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
759                         exporter = next;
760                         importer = vma;
761                         VM_WARN_ON(expand != importer);
762                 } else if (end < vma->vm_end) {
763                         /*
764                          * vma shrinks, and !insert tells it's not
765                          * split_vma inserting another: so it must be
766                          * mprotect case 4 shifting the boundary down.
767                          */
768                         adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
769                         exporter = vma;
770                         importer = next;
771                         VM_WARN_ON(expand != importer);
772                 }
773
774                 /*
775                  * Easily overlooked: when mprotect shifts the boundary,
776                  * make sure the expanding vma has anon_vma set if the
777                  * shrinking vma had, to cover any anon pages imported.
778                  */
779                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
780                         int error;
781
782                         importer->anon_vma = exporter->anon_vma;
783                         error = anon_vma_clone(importer, exporter);
784                         if (error)
785                                 return error;
786                 }
787         }
788 again:
789         vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
790
791         if (file) {
792                 mapping = file->f_mapping;
793                 root = &mapping->i_mmap;
794                 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
795
796                 if (adjust_next)
797                         uprobe_munmap(next, next->vm_start, next->vm_end);
798
799                 i_mmap_lock_write(mapping);
800                 if (insert) {
801                         /*
802                          * Put into interval tree now, so instantiated pages
803                          * are visible to arm/parisc __flush_dcache_page
804                          * throughout; but we cannot insert into address
805                          * space until vma start or end is updated.
806                          */
807                         __vma_link_file(insert);
808                 }
809         }
810
811         anon_vma = vma->anon_vma;
812         if (!anon_vma && adjust_next)
813                 anon_vma = next->anon_vma;
814         if (anon_vma) {
815                 VM_WARN_ON(adjust_next && next->anon_vma &&
816                            anon_vma != next->anon_vma);
817                 anon_vma_lock_write(anon_vma);
818                 anon_vma_interval_tree_pre_update_vma(vma);
819                 if (adjust_next)
820                         anon_vma_interval_tree_pre_update_vma(next);
821         }
822
823         if (root) {
824                 flush_dcache_mmap_lock(mapping);
825                 vma_interval_tree_remove(vma, root);
826                 if (adjust_next)
827                         vma_interval_tree_remove(next, root);
828         }
829
830         if (start != vma->vm_start) {
831                 vma->vm_start = start;
832                 start_changed = true;
833         }
834         if (end != vma->vm_end) {
835                 vma->vm_end = end;
836                 end_changed = true;
837         }
838         vma->vm_pgoff = pgoff;
839         if (adjust_next) {
840                 next->vm_start += adjust_next << PAGE_SHIFT;
841                 next->vm_pgoff += adjust_next;
842         }
843
844         if (root) {
845                 if (adjust_next)
846                         vma_interval_tree_insert(next, root);
847                 vma_interval_tree_insert(vma, root);
848                 flush_dcache_mmap_unlock(mapping);
849         }
850
851         if (remove_next) {
852                 /*
853                  * vma_merge has merged next into vma, and needs
854                  * us to remove next before dropping the locks.
855                  */
856                 if (remove_next != 3)
857                         __vma_unlink_prev(mm, next, vma);
858                 else
859                         /*
860                          * vma is not before next if they've been
861                          * swapped.
862                          *
863                          * pre-swap() next->vm_start was reduced so
864                          * tell validate_mm_rb to ignore pre-swap()
865                          * "next" (which is stored in post-swap()
866                          * "vma").
867                          */
868                         __vma_unlink_common(mm, next, NULL, false, vma);
869                 if (file)
870                         __remove_shared_vm_struct(next, file, mapping);
871         } else if (insert) {
872                 /*
873                  * split_vma has split insert from vma, and needs
874                  * us to insert it before dropping the locks
875                  * (it may either follow vma or precede it).
876                  */
877                 __insert_vm_struct(mm, insert);
878         } else {
879                 if (start_changed)
880                         vma_gap_update(vma);
881                 if (end_changed) {
882                         if (!next)
883                                 mm->highest_vm_end = vm_end_gap(vma);
884                         else if (!adjust_next)
885                                 vma_gap_update(next);
886                 }
887         }
888
889         if (anon_vma) {
890                 anon_vma_interval_tree_post_update_vma(vma);
891                 if (adjust_next)
892                         anon_vma_interval_tree_post_update_vma(next);
893                 anon_vma_unlock_write(anon_vma);
894         }
895         if (mapping)
896                 i_mmap_unlock_write(mapping);
897
898         if (root) {
899                 uprobe_mmap(vma);
900
901                 if (adjust_next)
902                         uprobe_mmap(next);
903         }
904
905         if (remove_next) {
906                 if (file) {
907                         uprobe_munmap(next, next->vm_start, next->vm_end);
908                         fput(file);
909                 }
910                 if (next->anon_vma)
911                         anon_vma_merge(vma, next);
912                 mm->map_count--;
913                 mpol_put(vma_policy(next));
914                 kmem_cache_free(vm_area_cachep, next);
915                 /*
916                  * In mprotect's case 6 (see comments on vma_merge),
917                  * we must remove another next too. It would clutter
918                  * up the code too much to do both in one go.
919                  */
920                 if (remove_next != 3) {
921                         /*
922                          * If "next" was removed and vma->vm_end was
923                          * expanded (up) over it, in turn
924                          * "next->vm_prev->vm_end" changed and the
925                          * "vma->vm_next" gap must be updated.
926                          */
927                         next = vma->vm_next;
928                 } else {
929                         /*
930                          * For the scope of the comment "next" and
931                          * "vma" considered pre-swap(): if "vma" was
932                          * removed, next->vm_start was expanded (down)
933                          * over it and the "next" gap must be updated.
934                          * Because of the swap() the post-swap() "vma"
935                          * actually points to pre-swap() "next"
936                          * (post-swap() "next" as opposed is now a
937                          * dangling pointer).
938                          */
939                         next = vma;
940                 }
941                 if (remove_next == 2) {
942                         remove_next = 1;
943                         end = next->vm_end;
944                         goto again;
945                 }
946                 else if (next)
947                         vma_gap_update(next);
948                 else {
949                         /*
950                          * If remove_next == 2 we obviously can't
951                          * reach this path.
952                          *
953                          * If remove_next == 3 we can't reach this
954                          * path because pre-swap() next is always not
955                          * NULL. pre-swap() "next" is not being
956                          * removed and its next->vm_end is not altered
957                          * (and furthermore "end" already matches
958                          * next->vm_end in remove_next == 3).
959                          *
960                          * We reach this only in the remove_next == 1
961                          * case if the "next" vma that was removed was
962                          * the highest vma of the mm. However in such
963                          * case next->vm_end == "end" and the extended
964                          * "vma" has vma->vm_end == next->vm_end so
965                          * mm->highest_vm_end doesn't need any update
966                          * in remove_next == 1 case.
967                          */
968                         VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
969                 }
970         }
971         if (insert && file)
972                 uprobe_mmap(insert);
973
974         validate_mm(mm);
975
976         return 0;
977 }
978
979 /*
980  * If the vma has a ->close operation then the driver probably needs to release
981  * per-vma resources, so we don't attempt to merge those.
982  */
983 static inline int is_mergeable_vma(struct vm_area_struct *vma,
984                                 struct file *file, unsigned long vm_flags,
985                                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
986 {
987         /*
988          * VM_SOFTDIRTY should not prevent from VMA merging, if we
989          * match the flags but dirty bit -- the caller should mark
990          * merged VMA as dirty. If dirty bit won't be excluded from
991          * comparison, we increase pressue on the memory system forcing
992          * the kernel to generate new VMAs when old one could be
993          * extended instead.
994          */
995         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
996                 return 0;
997         if (vma->vm_file != file)
998                 return 0;
999         if (vma->vm_ops && vma->vm_ops->close)
1000                 return 0;
1001         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1002                 return 0;
1003         return 1;
1004 }
1005
1006 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1007                                         struct anon_vma *anon_vma2,
1008                                         struct vm_area_struct *vma)
1009 {
1010         /*
1011          * The list_is_singular() test is to avoid merging VMA cloned from
1012          * parents. This can improve scalability caused by anon_vma lock.
1013          */
1014         if ((!anon_vma1 || !anon_vma2) && (!vma ||
1015                 list_is_singular(&vma->anon_vma_chain)))
1016                 return 1;
1017         return anon_vma1 == anon_vma2;
1018 }
1019
1020 /*
1021  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1022  * in front of (at a lower virtual address and file offset than) the vma.
1023  *
1024  * We cannot merge two vmas if they have differently assigned (non-NULL)
1025  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1026  *
1027  * We don't check here for the merged mmap wrapping around the end of pagecache
1028  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1029  * wrap, nor mmaps which cover the final page at index -1UL.
1030  */
1031 static int
1032 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1033                      struct anon_vma *anon_vma, struct file *file,
1034                      pgoff_t vm_pgoff,
1035                      struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1036 {
1037         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1038             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1039                 if (vma->vm_pgoff == vm_pgoff)
1040                         return 1;
1041         }
1042         return 0;
1043 }
1044
1045 /*
1046  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1047  * beyond (at a higher virtual address and file offset than) the vma.
1048  *
1049  * We cannot merge two vmas if they have differently assigned (non-NULL)
1050  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1051  */
1052 static int
1053 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1054                     struct anon_vma *anon_vma, struct file *file,
1055                     pgoff_t vm_pgoff,
1056                     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1057 {
1058         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1059             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1060                 pgoff_t vm_pglen;
1061                 vm_pglen = vma_pages(vma);
1062                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1063                         return 1;
1064         }
1065         return 0;
1066 }
1067
1068 /*
1069  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1070  * whether that can be merged with its predecessor or its successor.
1071  * Or both (it neatly fills a hole).
1072  *
1073  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1074  * certain not to be mapped by the time vma_merge is called; but when
1075  * called for mprotect, it is certain to be already mapped (either at
1076  * an offset within prev, or at the start of next), and the flags of
1077  * this area are about to be changed to vm_flags - and the no-change
1078  * case has already been eliminated.
1079  *
1080  * The following mprotect cases have to be considered, where AAAA is
1081  * the area passed down from mprotect_fixup, never extending beyond one
1082  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1083  *
1084  *     AAAA             AAAA                AAAA          AAAA
1085  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
1086  *    cannot merge    might become    might become    might become
1087  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
1088  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
1089  *    mremap move:                                    PPPPXXXXXXXX 8
1090  *        AAAA
1091  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
1092  *    might become    case 1 below    case 2 below    case 3 below
1093  *
1094  * It is important for case 8 that the the vma NNNN overlapping the
1095  * region AAAA is never going to extended over XXXX. Instead XXXX must
1096  * be extended in region AAAA and NNNN must be removed. This way in
1097  * all cases where vma_merge succeeds, the moment vma_adjust drops the
1098  * rmap_locks, the properties of the merged vma will be already
1099  * correct for the whole merged range. Some of those properties like
1100  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1101  * be correct for the whole merged range immediately after the
1102  * rmap_locks are released. Otherwise if XXXX would be removed and
1103  * NNNN would be extended over the XXXX range, remove_migration_ptes
1104  * or other rmap walkers (if working on addresses beyond the "end"
1105  * parameter) may establish ptes with the wrong permissions of NNNN
1106  * instead of the right permissions of XXXX.
1107  */
1108 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1109                         struct vm_area_struct *prev, unsigned long addr,
1110                         unsigned long end, unsigned long vm_flags,
1111                         struct anon_vma *anon_vma, struct file *file,
1112                         pgoff_t pgoff, struct mempolicy *policy,
1113                         struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1114 {
1115         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1116         struct vm_area_struct *area, *next;
1117         int err;
1118
1119         /*
1120          * We later require that vma->vm_flags == vm_flags,
1121          * so this tests vma->vm_flags & VM_SPECIAL, too.
1122          */
1123         if (vm_flags & VM_SPECIAL)
1124                 return NULL;
1125
1126         if (prev)
1127                 next = prev->vm_next;
1128         else
1129                 next = mm->mmap;
1130         area = next;
1131         if (area && area->vm_end == end)                /* cases 6, 7, 8 */
1132                 next = next->vm_next;
1133
1134         /* verify some invariant that must be enforced by the caller */
1135         VM_WARN_ON(prev && addr <= prev->vm_start);
1136         VM_WARN_ON(area && end > area->vm_end);
1137         VM_WARN_ON(addr >= end);
1138
1139         /*
1140          * Can it merge with the predecessor?
1141          */
1142         if (prev && prev->vm_end == addr &&
1143                         mpol_equal(vma_policy(prev), policy) &&
1144                         can_vma_merge_after(prev, vm_flags,
1145                                             anon_vma, file, pgoff,
1146                                             vm_userfaultfd_ctx)) {
1147                 /*
1148                  * OK, it can.  Can we now merge in the successor as well?
1149                  */
1150                 if (next && end == next->vm_start &&
1151                                 mpol_equal(policy, vma_policy(next)) &&
1152                                 can_vma_merge_before(next, vm_flags,
1153                                                      anon_vma, file,
1154                                                      pgoff+pglen,
1155                                                      vm_userfaultfd_ctx) &&
1156                                 is_mergeable_anon_vma(prev->anon_vma,
1157                                                       next->anon_vma, NULL)) {
1158                                                         /* cases 1, 6 */
1159                         err = __vma_adjust(prev, prev->vm_start,
1160                                          next->vm_end, prev->vm_pgoff, NULL,
1161                                          prev);
1162                 } else                                  /* cases 2, 5, 7 */
1163                         err = __vma_adjust(prev, prev->vm_start,
1164                                          end, prev->vm_pgoff, NULL, prev);
1165                 if (err)
1166                         return NULL;
1167                 khugepaged_enter_vma_merge(prev, vm_flags);
1168                 return prev;
1169         }
1170
1171         /*
1172          * Can this new request be merged in front of next?
1173          */
1174         if (next && end == next->vm_start &&
1175                         mpol_equal(policy, vma_policy(next)) &&
1176                         can_vma_merge_before(next, vm_flags,
1177                                              anon_vma, file, pgoff+pglen,
1178                                              vm_userfaultfd_ctx)) {
1179                 if (prev && addr < prev->vm_end)        /* case 4 */
1180                         err = __vma_adjust(prev, prev->vm_start,
1181                                          addr, prev->vm_pgoff, NULL, next);
1182                 else {                                  /* cases 3, 8 */
1183                         err = __vma_adjust(area, addr, next->vm_end,
1184                                          next->vm_pgoff - pglen, NULL, next);
1185                         /*
1186                          * In case 3 area is already equal to next and
1187                          * this is a noop, but in case 8 "area" has
1188                          * been removed and next was expanded over it.
1189                          */
1190                         area = next;
1191                 }
1192                 if (err)
1193                         return NULL;
1194                 khugepaged_enter_vma_merge(area, vm_flags);
1195                 return area;
1196         }
1197
1198         return NULL;
1199 }
1200
1201 /*
1202  * Rough compatbility check to quickly see if it's even worth looking
1203  * at sharing an anon_vma.
1204  *
1205  * They need to have the same vm_file, and the flags can only differ
1206  * in things that mprotect may change.
1207  *
1208  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1209  * we can merge the two vma's. For example, we refuse to merge a vma if
1210  * there is a vm_ops->close() function, because that indicates that the
1211  * driver is doing some kind of reference counting. But that doesn't
1212  * really matter for the anon_vma sharing case.
1213  */
1214 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1215 {
1216         return a->vm_end == b->vm_start &&
1217                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1218                 a->vm_file == b->vm_file &&
1219                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1220                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1221 }
1222
1223 /*
1224  * Do some basic sanity checking to see if we can re-use the anon_vma
1225  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1226  * the same as 'old', the other will be the new one that is trying
1227  * to share the anon_vma.
1228  *
1229  * NOTE! This runs with mm_sem held for reading, so it is possible that
1230  * the anon_vma of 'old' is concurrently in the process of being set up
1231  * by another page fault trying to merge _that_. But that's ok: if it
1232  * is being set up, that automatically means that it will be a singleton
1233  * acceptable for merging, so we can do all of this optimistically. But
1234  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1235  *
1236  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1237  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1238  * is to return an anon_vma that is "complex" due to having gone through
1239  * a fork).
1240  *
1241  * We also make sure that the two vma's are compatible (adjacent,
1242  * and with the same memory policies). That's all stable, even with just
1243  * a read lock on the mm_sem.
1244  */
1245 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1246 {
1247         if (anon_vma_compatible(a, b)) {
1248                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1249
1250                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1251                         return anon_vma;
1252         }
1253         return NULL;
1254 }
1255
1256 /*
1257  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1258  * neighbouring vmas for a suitable anon_vma, before it goes off
1259  * to allocate a new anon_vma.  It checks because a repetitive
1260  * sequence of mprotects and faults may otherwise lead to distinct
1261  * anon_vmas being allocated, preventing vma merge in subsequent
1262  * mprotect.
1263  */
1264 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1265 {
1266         struct anon_vma *anon_vma;
1267         struct vm_area_struct *near;
1268
1269         near = vma->vm_next;
1270         if (!near)
1271                 goto try_prev;
1272
1273         anon_vma = reusable_anon_vma(near, vma, near);
1274         if (anon_vma)
1275                 return anon_vma;
1276 try_prev:
1277         near = vma->vm_prev;
1278         if (!near)
1279                 goto none;
1280
1281         anon_vma = reusable_anon_vma(near, near, vma);
1282         if (anon_vma)
1283                 return anon_vma;
1284 none:
1285         /*
1286          * There's no absolute need to look only at touching neighbours:
1287          * we could search further afield for "compatible" anon_vmas.
1288          * But it would probably just be a waste of time searching,
1289          * or lead to too many vmas hanging off the same anon_vma.
1290          * We're trying to allow mprotect remerging later on,
1291          * not trying to minimize memory used for anon_vmas.
1292          */
1293         return NULL;
1294 }
1295
1296 /*
1297  * If a hint addr is less than mmap_min_addr change hint to be as
1298  * low as possible but still greater than mmap_min_addr
1299  */
1300 static inline unsigned long round_hint_to_min(unsigned long hint)
1301 {
1302         hint &= PAGE_MASK;
1303         if (((void *)hint != NULL) &&
1304             (hint < mmap_min_addr))
1305                 return PAGE_ALIGN(mmap_min_addr);
1306         return hint;
1307 }
1308
1309 static inline int mlock_future_check(struct mm_struct *mm,
1310                                      unsigned long flags,
1311                                      unsigned long len)
1312 {
1313         unsigned long locked, lock_limit;
1314
1315         /*  mlock MCL_FUTURE? */
1316         if (flags & VM_LOCKED) {
1317                 locked = len >> PAGE_SHIFT;
1318                 locked += mm->locked_vm;
1319                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1320                 lock_limit >>= PAGE_SHIFT;
1321                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1322                         return -EAGAIN;
1323         }
1324         return 0;
1325 }
1326
1327 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1328 {
1329         if (S_ISREG(inode->i_mode))
1330                 return MAX_LFS_FILESIZE;
1331
1332         if (S_ISBLK(inode->i_mode))
1333                 return MAX_LFS_FILESIZE;
1334
1335         /* Special "we do even unsigned file positions" case */
1336         if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1337                 return 0;
1338
1339         /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1340         return ULONG_MAX;
1341 }
1342
1343 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1344                                 unsigned long pgoff, unsigned long len)
1345 {
1346         u64 maxsize = file_mmap_size_max(file, inode);
1347
1348         if (maxsize && len > maxsize)
1349                 return false;
1350         maxsize -= len;
1351         if (pgoff > maxsize >> PAGE_SHIFT)
1352                 return false;
1353         return true;
1354 }
1355
1356 /*
1357  * The caller must hold down_write(&current->mm->mmap_sem).
1358  */
1359 unsigned long do_mmap(struct file *file, unsigned long addr,
1360                         unsigned long len, unsigned long prot,
1361                         unsigned long flags, vm_flags_t vm_flags,
1362                         unsigned long pgoff, unsigned long *populate,
1363                         struct list_head *uf)
1364 {
1365         struct mm_struct *mm = current->mm;
1366         int pkey = 0;
1367
1368         *populate = 0;
1369
1370         if (!len)
1371                 return -EINVAL;
1372
1373         /*
1374          * Does the application expect PROT_READ to imply PROT_EXEC?
1375          *
1376          * (the exception is when the underlying filesystem is noexec
1377          *  mounted, in which case we dont add PROT_EXEC.)
1378          */
1379         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1380                 if (!(file && path_noexec(&file->f_path)))
1381                         prot |= PROT_EXEC;
1382
1383         /* force arch specific MAP_FIXED handling in get_unmapped_area */
1384         if (flags & MAP_FIXED_NOREPLACE)
1385                 flags |= MAP_FIXED;
1386
1387         if (!(flags & MAP_FIXED))
1388                 addr = round_hint_to_min(addr);
1389
1390         /* Careful about overflows.. */
1391         len = PAGE_ALIGN(len);
1392         if (!len)
1393                 return -ENOMEM;
1394
1395         /* offset overflow? */
1396         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1397                 return -EOVERFLOW;
1398
1399         /* Too many mappings? */
1400         if (mm->map_count > sysctl_max_map_count)
1401                 return -ENOMEM;
1402
1403         /* Obtain the address to map to. we verify (or select) it and ensure
1404          * that it represents a valid section of the address space.
1405          */
1406         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1407         if (offset_in_page(addr))
1408                 return addr;
1409
1410         if (flags & MAP_FIXED_NOREPLACE) {
1411                 struct vm_area_struct *vma = find_vma(mm, addr);
1412
1413                 if (vma && vma->vm_start <= addr)
1414                         return -EEXIST;
1415         }
1416
1417         if (prot == PROT_EXEC) {
1418                 pkey = execute_only_pkey(mm);
1419                 if (pkey < 0)
1420                         pkey = 0;
1421         }
1422
1423         /* Do simple checking here so the lower-level routines won't have
1424          * to. we assume access permissions have been handled by the open
1425          * of the memory object, so we don't do any here.
1426          */
1427         vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1428                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1429
1430         if (flags & MAP_LOCKED)
1431                 if (!can_do_mlock())
1432                         return -EPERM;
1433
1434         if (mlock_future_check(mm, vm_flags, len))
1435                 return -EAGAIN;
1436
1437         if (file) {
1438                 struct inode *inode = file_inode(file);
1439                 unsigned long flags_mask;
1440
1441                 if (!file_mmap_ok(file, inode, pgoff, len))
1442                         return -EOVERFLOW;
1443
1444                 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1445
1446                 switch (flags & MAP_TYPE) {
1447                 case MAP_SHARED:
1448                         /*
1449                          * Force use of MAP_SHARED_VALIDATE with non-legacy
1450                          * flags. E.g. MAP_SYNC is dangerous to use with
1451                          * MAP_SHARED as you don't know which consistency model
1452                          * you will get. We silently ignore unsupported flags
1453                          * with MAP_SHARED to preserve backward compatibility.
1454                          */
1455                         flags &= LEGACY_MAP_MASK;
1456                         /* fall through */
1457                 case MAP_SHARED_VALIDATE:
1458                         if (flags & ~flags_mask)
1459                                 return -EOPNOTSUPP;
1460                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1461                                 return -EACCES;
1462
1463                         /*
1464                          * Make sure we don't allow writing to an append-only
1465                          * file..
1466                          */
1467                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1468                                 return -EACCES;
1469
1470                         /*
1471                          * Make sure there are no mandatory locks on the file.
1472                          */
1473                         if (locks_verify_locked(file))
1474                                 return -EAGAIN;
1475
1476                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1477                         if (!(file->f_mode & FMODE_WRITE))
1478                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1479
1480                         /* fall through */
1481                 case MAP_PRIVATE:
1482                         if (!(file->f_mode & FMODE_READ))
1483                                 return -EACCES;
1484                         if (path_noexec(&file->f_path)) {
1485                                 if (vm_flags & VM_EXEC)
1486                                         return -EPERM;
1487                                 vm_flags &= ~VM_MAYEXEC;
1488                         }
1489
1490                         if (!file->f_op->mmap)
1491                                 return -ENODEV;
1492                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1493                                 return -EINVAL;
1494                         break;
1495
1496                 default:
1497                         return -EINVAL;
1498                 }
1499         } else {
1500                 switch (flags & MAP_TYPE) {
1501                 case MAP_SHARED:
1502                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1503                                 return -EINVAL;
1504                         /*
1505                          * Ignore pgoff.
1506                          */
1507                         pgoff = 0;
1508                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1509                         break;
1510                 case MAP_PRIVATE:
1511                         /*
1512                          * Set pgoff according to addr for anon_vma.
1513                          */
1514                         pgoff = addr >> PAGE_SHIFT;
1515                         break;
1516                 default:
1517                         return -EINVAL;
1518                 }
1519         }
1520
1521         /*
1522          * Set 'VM_NORESERVE' if we should not account for the
1523          * memory use of this mapping.
1524          */
1525         if (flags & MAP_NORESERVE) {
1526                 /* We honor MAP_NORESERVE if allowed to overcommit */
1527                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1528                         vm_flags |= VM_NORESERVE;
1529
1530                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1531                 if (file && is_file_hugepages(file))
1532                         vm_flags |= VM_NORESERVE;
1533         }
1534
1535         addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1536         if (!IS_ERR_VALUE(addr) &&
1537             ((vm_flags & VM_LOCKED) ||
1538              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1539                 *populate = len;
1540         return addr;
1541 }
1542
1543 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1544                               unsigned long prot, unsigned long flags,
1545                               unsigned long fd, unsigned long pgoff)
1546 {
1547         struct file *file = NULL;
1548         unsigned long retval;
1549
1550         if (!(flags & MAP_ANONYMOUS)) {
1551                 audit_mmap_fd(fd, flags);
1552                 file = fget(fd);
1553                 if (!file)
1554                         return -EBADF;
1555                 if (is_file_hugepages(file))
1556                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1557                 retval = -EINVAL;
1558                 if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1559                         goto out_fput;
1560         } else if (flags & MAP_HUGETLB) {
1561                 struct user_struct *user = NULL;
1562                 struct hstate *hs;
1563
1564                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1565                 if (!hs)
1566                         return -EINVAL;
1567
1568                 len = ALIGN(len, huge_page_size(hs));
1569                 /*
1570                  * VM_NORESERVE is used because the reservations will be
1571                  * taken when vm_ops->mmap() is called
1572                  * A dummy user value is used because we are not locking
1573                  * memory so no accounting is necessary
1574                  */
1575                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1576                                 VM_NORESERVE,
1577                                 &user, HUGETLB_ANONHUGE_INODE,
1578                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1579                 if (IS_ERR(file))
1580                         return PTR_ERR(file);
1581         }
1582
1583         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1584
1585         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1586 out_fput:
1587         if (file)
1588                 fput(file);
1589         return retval;
1590 }
1591
1592 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1593                 unsigned long, prot, unsigned long, flags,
1594                 unsigned long, fd, unsigned long, pgoff)
1595 {
1596         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1597 }
1598
1599 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1600 struct mmap_arg_struct {
1601         unsigned long addr;
1602         unsigned long len;
1603         unsigned long prot;
1604         unsigned long flags;
1605         unsigned long fd;
1606         unsigned long offset;
1607 };
1608
1609 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1610 {
1611         struct mmap_arg_struct a;
1612
1613         if (copy_from_user(&a, arg, sizeof(a)))
1614                 return -EFAULT;
1615         if (offset_in_page(a.offset))
1616                 return -EINVAL;
1617
1618         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1619                                a.offset >> PAGE_SHIFT);
1620 }
1621 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1622
1623 /*
1624  * Some shared mappigns will want the pages marked read-only
1625  * to track write events. If so, we'll downgrade vm_page_prot
1626  * to the private version (using protection_map[] without the
1627  * VM_SHARED bit).
1628  */
1629 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1630 {
1631         vm_flags_t vm_flags = vma->vm_flags;
1632         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1633
1634         /* If it was private or non-writable, the write bit is already clear */
1635         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1636                 return 0;
1637
1638         /* The backer wishes to know when pages are first written to? */
1639         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1640                 return 1;
1641
1642         /* The open routine did something to the protections that pgprot_modify
1643          * won't preserve? */
1644         if (pgprot_val(vm_page_prot) !=
1645             pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1646                 return 0;
1647
1648         /* Do we need to track softdirty? */
1649         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1650                 return 1;
1651
1652         /* Specialty mapping? */
1653         if (vm_flags & VM_PFNMAP)
1654                 return 0;
1655
1656         /* Can the mapping track the dirty pages? */
1657         return vma->vm_file && vma->vm_file->f_mapping &&
1658                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1659 }
1660
1661 /*
1662  * We account for memory if it's a private writeable mapping,
1663  * not hugepages and VM_NORESERVE wasn't set.
1664  */
1665 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1666 {
1667         /*
1668          * hugetlb has its own accounting separate from the core VM
1669          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1670          */
1671         if (file && is_file_hugepages(file))
1672                 return 0;
1673
1674         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1675 }
1676
1677 unsigned long mmap_region(struct file *file, unsigned long addr,
1678                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1679                 struct list_head *uf)
1680 {
1681         struct mm_struct *mm = current->mm;
1682         struct vm_area_struct *vma, *prev;
1683         int error;
1684         struct rb_node **rb_link, *rb_parent;
1685         unsigned long charged = 0;
1686
1687         /* Check against address space limit. */
1688         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1689                 unsigned long nr_pages;
1690
1691                 /*
1692                  * MAP_FIXED may remove pages of mappings that intersects with
1693                  * requested mapping. Account for the pages it would unmap.
1694                  */
1695                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1696
1697                 if (!may_expand_vm(mm, vm_flags,
1698                                         (len >> PAGE_SHIFT) - nr_pages))
1699                         return -ENOMEM;
1700         }
1701
1702         /* Clear old maps */
1703         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1704                               &rb_parent)) {
1705                 if (do_munmap(mm, addr, len, uf))
1706                         return -ENOMEM;
1707         }
1708
1709         /*
1710          * Private writable mapping: check memory availability
1711          */
1712         if (accountable_mapping(file, vm_flags)) {
1713                 charged = len >> PAGE_SHIFT;
1714                 if (security_vm_enough_memory_mm(mm, charged))
1715                         return -ENOMEM;
1716                 vm_flags |= VM_ACCOUNT;
1717         }
1718
1719         /*
1720          * Can we just expand an old mapping?
1721          */
1722         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1723                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1724         if (vma)
1725                 goto out;
1726
1727         /*
1728          * Determine the object being mapped and call the appropriate
1729          * specific mapper. the address has already been validated, but
1730          * not unmapped, but the maps are removed from the list.
1731          */
1732         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1733         if (!vma) {
1734                 error = -ENOMEM;
1735                 goto unacct_error;
1736         }
1737
1738         vma->vm_mm = mm;
1739         vma->vm_start = addr;
1740         vma->vm_end = addr + len;
1741         vma->vm_flags = vm_flags;
1742         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1743         vma->vm_pgoff = pgoff;
1744         INIT_LIST_HEAD(&vma->anon_vma_chain);
1745
1746         if (file) {
1747                 if (vm_flags & VM_DENYWRITE) {
1748                         error = deny_write_access(file);
1749                         if (error)
1750                                 goto free_vma;
1751                 }
1752                 if (vm_flags & VM_SHARED) {
1753                         error = mapping_map_writable(file->f_mapping);
1754                         if (error)
1755                                 goto allow_write_and_free_vma;
1756                 }
1757
1758                 /* ->mmap() can change vma->vm_file, but must guarantee that
1759                  * vma_link() below can deny write-access if VM_DENYWRITE is set
1760                  * and map writably if VM_SHARED is set. This usually means the
1761                  * new file must not have been exposed to user-space, yet.
1762                  */
1763                 vma->vm_file = get_file(file);
1764                 error = call_mmap(file, vma);
1765                 if (error)
1766                         goto unmap_and_free_vma;
1767
1768                 /* Can addr have changed??
1769                  *
1770                  * Answer: Yes, several device drivers can do it in their
1771                  *         f_op->mmap method. -DaveM
1772                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1773                  *      be updated for vma_link()
1774                  */
1775                 WARN_ON_ONCE(addr != vma->vm_start);
1776
1777                 addr = vma->vm_start;
1778                 vm_flags = vma->vm_flags;
1779         } else if (vm_flags & VM_SHARED) {
1780                 error = shmem_zero_setup(vma);
1781                 if (error)
1782                         goto free_vma;
1783         }
1784
1785         vma_link(mm, vma, prev, rb_link, rb_parent);
1786         /* Once vma denies write, undo our temporary denial count */
1787         if (file) {
1788                 if (vm_flags & VM_SHARED)
1789                         mapping_unmap_writable(file->f_mapping);
1790                 if (vm_flags & VM_DENYWRITE)
1791                         allow_write_access(file);
1792         }
1793         file = vma->vm_file;
1794 out:
1795         perf_event_mmap(vma);
1796
1797         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1798         if (vm_flags & VM_LOCKED) {
1799                 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1800                                         vma == get_gate_vma(current->mm)))
1801                         mm->locked_vm += (len >> PAGE_SHIFT);
1802                 else
1803                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1804         }
1805
1806         if (file)
1807                 uprobe_mmap(vma);
1808
1809         /*
1810          * New (or expanded) vma always get soft dirty status.
1811          * Otherwise user-space soft-dirty page tracker won't
1812          * be able to distinguish situation when vma area unmapped,
1813          * then new mapped in-place (which must be aimed as
1814          * a completely new data area).
1815          */
1816         vma->vm_flags |= VM_SOFTDIRTY;
1817
1818         vma_set_page_prot(vma);
1819
1820         return addr;
1821
1822 unmap_and_free_vma:
1823         vma->vm_file = NULL;
1824         fput(file);
1825
1826         /* Undo any partial mapping done by a device driver. */
1827         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1828         charged = 0;
1829         if (vm_flags & VM_SHARED)
1830                 mapping_unmap_writable(file->f_mapping);
1831 allow_write_and_free_vma:
1832         if (vm_flags & VM_DENYWRITE)
1833                 allow_write_access(file);
1834 free_vma:
1835         kmem_cache_free(vm_area_cachep, vma);
1836 unacct_error:
1837         if (charged)
1838                 vm_unacct_memory(charged);
1839         return error;
1840 }
1841
1842 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1843 {
1844         /*
1845          * We implement the search by looking for an rbtree node that
1846          * immediately follows a suitable gap. That is,
1847          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1848          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1849          * - gap_end - gap_start >= length
1850          */
1851
1852         struct mm_struct *mm = current->mm;
1853         struct vm_area_struct *vma;
1854         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1855
1856         /* Adjust search length to account for worst case alignment overhead */
1857         length = info->length + info->align_mask;
1858         if (length < info->length)
1859                 return -ENOMEM;
1860
1861         /* Adjust search limits by the desired length */
1862         if (info->high_limit < length)
1863                 return -ENOMEM;
1864         high_limit = info->high_limit - length;
1865
1866         if (info->low_limit > high_limit)
1867                 return -ENOMEM;
1868         low_limit = info->low_limit + length;
1869
1870         /* Check if rbtree root looks promising */
1871         if (RB_EMPTY_ROOT(&mm->mm_rb))
1872                 goto check_highest;
1873         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1874         if (vma->rb_subtree_gap < length)
1875                 goto check_highest;
1876
1877         while (true) {
1878                 /* Visit left subtree if it looks promising */
1879                 gap_end = vm_start_gap(vma);
1880                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1881                         struct vm_area_struct *left =
1882                                 rb_entry(vma->vm_rb.rb_left,
1883                                          struct vm_area_struct, vm_rb);
1884                         if (left->rb_subtree_gap >= length) {
1885                                 vma = left;
1886                                 continue;
1887                         }
1888                 }
1889
1890                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1891 check_current:
1892                 /* Check if current node has a suitable gap */
1893                 if (gap_start > high_limit)
1894                         return -ENOMEM;
1895                 if (gap_end >= low_limit &&
1896                     gap_end > gap_start && gap_end - gap_start >= length)
1897                         goto found;
1898
1899                 /* Visit right subtree if it looks promising */
1900                 if (vma->vm_rb.rb_right) {
1901                         struct vm_area_struct *right =
1902                                 rb_entry(vma->vm_rb.rb_right,
1903                                          struct vm_area_struct, vm_rb);
1904                         if (right->rb_subtree_gap >= length) {
1905                                 vma = right;
1906                                 continue;
1907                         }
1908                 }
1909
1910                 /* Go back up the rbtree to find next candidate node */
1911                 while (true) {
1912                         struct rb_node *prev = &vma->vm_rb;
1913                         if (!rb_parent(prev))
1914                                 goto check_highest;
1915                         vma = rb_entry(rb_parent(prev),
1916                                        struct vm_area_struct, vm_rb);
1917                         if (prev == vma->vm_rb.rb_left) {
1918                                 gap_start = vm_end_gap(vma->vm_prev);
1919                                 gap_end = vm_start_gap(vma);
1920                                 goto check_current;
1921                         }
1922                 }
1923         }
1924
1925 check_highest:
1926         /* Check highest gap, which does not precede any rbtree node */
1927         gap_start = mm->highest_vm_end;
1928         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1929         if (gap_start > high_limit)
1930                 return -ENOMEM;
1931
1932 found:
1933         /* We found a suitable gap. Clip it with the original low_limit. */
1934         if (gap_start < info->low_limit)
1935                 gap_start = info->low_limit;
1936
1937         /* Adjust gap address to the desired alignment */
1938         gap_start += (info->align_offset - gap_start) & info->align_mask;
1939
1940         VM_BUG_ON(gap_start + info->length > info->high_limit);
1941         VM_BUG_ON(gap_start + info->length > gap_end);
1942         return gap_start;
1943 }
1944
1945 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1946 {
1947         struct mm_struct *mm = current->mm;
1948         struct vm_area_struct *vma;
1949         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1950
1951         /* Adjust search length to account for worst case alignment overhead */
1952         length = info->length + info->align_mask;
1953         if (length < info->length)
1954                 return -ENOMEM;
1955
1956         /*
1957          * Adjust search limits by the desired length.
1958          * See implementation comment at top of unmapped_area().
1959          */
1960         gap_end = info->high_limit;
1961         if (gap_end < length)
1962                 return -ENOMEM;
1963         high_limit = gap_end - length;
1964
1965         if (info->low_limit > high_limit)
1966                 return -ENOMEM;
1967         low_limit = info->low_limit + length;
1968
1969         /* Check highest gap, which does not precede any rbtree node */
1970         gap_start = mm->highest_vm_end;
1971         if (gap_start <= high_limit)
1972                 goto found_highest;
1973
1974         /* Check if rbtree root looks promising */
1975         if (RB_EMPTY_ROOT(&mm->mm_rb))
1976                 return -ENOMEM;
1977         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1978         if (vma->rb_subtree_gap < length)
1979                 return -ENOMEM;
1980
1981         while (true) {
1982                 /* Visit right subtree if it looks promising */
1983                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1984                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1985                         struct vm_area_struct *right =
1986                                 rb_entry(vma->vm_rb.rb_right,
1987                                          struct vm_area_struct, vm_rb);
1988                         if (right->rb_subtree_gap >= length) {
1989                                 vma = right;
1990                                 continue;
1991                         }
1992                 }
1993
1994 check_current:
1995                 /* Check if current node has a suitable gap */
1996                 gap_end = vm_start_gap(vma);
1997                 if (gap_end < low_limit)
1998                         return -ENOMEM;
1999                 if (gap_start <= high_limit &&
2000                     gap_end > gap_start && gap_end - gap_start >= length)
2001                         goto found;
2002
2003                 /* Visit left subtree if it looks promising */
2004                 if (vma->vm_rb.rb_left) {
2005                         struct vm_area_struct *left =
2006                                 rb_entry(vma->vm_rb.rb_left,
2007                                          struct vm_area_struct, vm_rb);
2008                         if (left->rb_subtree_gap >= length) {
2009                                 vma = left;
2010                                 continue;
2011                         }
2012                 }
2013
2014                 /* Go back up the rbtree to find next candidate node */
2015                 while (true) {
2016                         struct rb_node *prev = &vma->vm_rb;
2017                         if (!rb_parent(prev))
2018                                 return -ENOMEM;
2019                         vma = rb_entry(rb_parent(prev),
2020                                        struct vm_area_struct, vm_rb);
2021                         if (prev == vma->vm_rb.rb_right) {
2022                                 gap_start = vma->vm_prev ?
2023                                         vm_end_gap(vma->vm_prev) : 0;
2024                                 goto check_current;
2025                         }
2026                 }
2027         }
2028
2029 found:
2030         /* We found a suitable gap. Clip it with the original high_limit. */
2031         if (gap_end > info->high_limit)
2032                 gap_end = info->high_limit;
2033
2034 found_highest:
2035         /* Compute highest gap address at the desired alignment */
2036         gap_end -= info->length;
2037         gap_end -= (gap_end - info->align_offset) & info->align_mask;
2038
2039         VM_BUG_ON(gap_end < info->low_limit);
2040         VM_BUG_ON(gap_end < gap_start);
2041         return gap_end;
2042 }
2043
2044 /* Get an address range which is currently unmapped.
2045  * For shmat() with addr=0.
2046  *
2047  * Ugly calling convention alert:
2048  * Return value with the low bits set means error value,
2049  * ie
2050  *      if (ret & ~PAGE_MASK)
2051  *              error = ret;
2052  *
2053  * This function "knows" that -ENOMEM has the bits set.
2054  */
2055 #ifndef HAVE_ARCH_UNMAPPED_AREA
2056 unsigned long
2057 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2058                 unsigned long len, unsigned long pgoff, unsigned long flags)
2059 {
2060         struct mm_struct *mm = current->mm;
2061         struct vm_area_struct *vma, *prev;
2062         struct vm_unmapped_area_info info;
2063
2064         if (len > TASK_SIZE - mmap_min_addr)
2065                 return -ENOMEM;
2066
2067         if (flags & MAP_FIXED)
2068                 return addr;
2069
2070         if (addr) {
2071                 addr = PAGE_ALIGN(addr);
2072                 vma = find_vma_prev(mm, addr, &prev);
2073                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2074                     (!vma || addr + len <= vm_start_gap(vma)) &&
2075                     (!prev || addr >= vm_end_gap(prev)))
2076                         return addr;
2077         }
2078
2079         info.flags = 0;
2080         info.length = len;
2081         info.low_limit = mm->mmap_base;
2082         info.high_limit = TASK_SIZE;
2083         info.align_mask = 0;
2084         return vm_unmapped_area(&info);
2085 }
2086 #endif
2087
2088 /*
2089  * This mmap-allocator allocates new areas top-down from below the
2090  * stack's low limit (the base):
2091  */
2092 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2093 unsigned long
2094 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2095                           const unsigned long len, const unsigned long pgoff,
2096                           const unsigned long flags)
2097 {
2098         struct vm_area_struct *vma, *prev;
2099         struct mm_struct *mm = current->mm;
2100         unsigned long addr = addr0;
2101         struct vm_unmapped_area_info info;
2102
2103         /* requested length too big for entire address space */
2104         if (len > TASK_SIZE - mmap_min_addr)
2105                 return -ENOMEM;
2106
2107         if (flags & MAP_FIXED)
2108                 return addr;
2109
2110         /* requesting a specific address */
2111         if (addr) {
2112                 addr = PAGE_ALIGN(addr);
2113                 vma = find_vma_prev(mm, addr, &prev);
2114                 if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2115                                 (!vma || addr + len <= vm_start_gap(vma)) &&
2116                                 (!prev || addr >= vm_end_gap(prev)))
2117                         return addr;
2118         }
2119
2120         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2121         info.length = len;
2122         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2123         info.high_limit = mm->mmap_base;
2124         info.align_mask = 0;
2125         addr = vm_unmapped_area(&info);
2126
2127         /*
2128          * A failed mmap() very likely causes application failure,
2129          * so fall back to the bottom-up function here. This scenario
2130          * can happen with large stack limits and large mmap()
2131          * allocations.
2132          */
2133         if (offset_in_page(addr)) {
2134                 VM_BUG_ON(addr != -ENOMEM);
2135                 info.flags = 0;
2136                 info.low_limit = TASK_UNMAPPED_BASE;
2137                 info.high_limit = TASK_SIZE;
2138                 addr = vm_unmapped_area(&info);
2139         }
2140
2141         return addr;
2142 }
2143 #endif
2144
2145 unsigned long
2146 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2147                 unsigned long pgoff, unsigned long flags)
2148 {
2149         unsigned long (*get_area)(struct file *, unsigned long,
2150                                   unsigned long, unsigned long, unsigned long);
2151
2152         unsigned long error = arch_mmap_check(addr, len, flags);
2153         if (error)
2154                 return error;
2155
2156         /* Careful about overflows.. */
2157         if (len > TASK_SIZE)
2158                 return -ENOMEM;
2159
2160         get_area = current->mm->get_unmapped_area;
2161         if (file) {
2162                 if (file->f_op->get_unmapped_area)
2163                         get_area = file->f_op->get_unmapped_area;
2164         } else if (flags & MAP_SHARED) {
2165                 /*
2166                  * mmap_region() will call shmem_zero_setup() to create a file,
2167                  * so use shmem's get_unmapped_area in case it can be huge.
2168                  * do_mmap_pgoff() will clear pgoff, so match alignment.
2169                  */
2170                 pgoff = 0;
2171                 get_area = shmem_get_unmapped_area;
2172         }
2173
2174         addr = get_area(file, addr, len, pgoff, flags);
2175         if (IS_ERR_VALUE(addr))
2176                 return addr;
2177
2178         if (addr > TASK_SIZE - len)
2179                 return -ENOMEM;
2180         if (offset_in_page(addr))
2181                 return -EINVAL;
2182
2183         error = security_mmap_addr(addr);
2184         return error ? error : addr;
2185 }
2186
2187 EXPORT_SYMBOL(get_unmapped_area);
2188
2189 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2190 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2191 {
2192         struct rb_node *rb_node;
2193         struct vm_area_struct *vma;
2194
2195         /* Check the cache first. */
2196         vma = vmacache_find(mm, addr);
2197         if (likely(vma))
2198                 return vma;
2199
2200         rb_node = mm->mm_rb.rb_node;
2201
2202         while (rb_node) {
2203                 struct vm_area_struct *tmp;
2204
2205                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2206
2207                 if (tmp->vm_end > addr) {
2208                         vma = tmp;
2209                         if (tmp->vm_start <= addr)
2210                                 break;
2211                         rb_node = rb_node->rb_left;
2212                 } else
2213                         rb_node = rb_node->rb_right;
2214         }
2215
2216         if (vma)
2217                 vmacache_update(addr, vma);
2218         return vma;
2219 }
2220
2221 EXPORT_SYMBOL(find_vma);
2222
2223 /*
2224  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2225  */
2226 struct vm_area_struct *
2227 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2228                         struct vm_area_struct **pprev)
2229 {
2230         struct vm_area_struct *vma;
2231
2232         vma = find_vma(mm, addr);
2233         if (vma) {
2234                 *pprev = vma->vm_prev;
2235         } else {
2236                 struct rb_node *rb_node = mm->mm_rb.rb_node;
2237                 *pprev = NULL;
2238                 while (rb_node) {
2239                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2240                         rb_node = rb_node->rb_right;
2241                 }
2242         }
2243         return vma;
2244 }
2245
2246 /*
2247  * Verify that the stack growth is acceptable and
2248  * update accounting. This is shared with both the
2249  * grow-up and grow-down cases.
2250  */
2251 static int acct_stack_growth(struct vm_area_struct *vma,
2252                              unsigned long size, unsigned long grow)
2253 {
2254         struct mm_struct *mm = vma->vm_mm;
2255         unsigned long new_start;
2256
2257         /* address space limit tests */
2258         if (!may_expand_vm(mm, vma->vm_flags, grow))
2259                 return -ENOMEM;
2260
2261         /* Stack limit test */
2262         if (size > rlimit(RLIMIT_STACK))
2263                 return -ENOMEM;
2264
2265         /* mlock limit tests */
2266         if (vma->vm_flags & VM_LOCKED) {
2267                 unsigned long locked;
2268                 unsigned long limit;
2269                 locked = mm->locked_vm + grow;
2270                 limit = rlimit(RLIMIT_MEMLOCK);
2271                 limit >>= PAGE_SHIFT;
2272                 if (locked > limit && !capable(CAP_IPC_LOCK))
2273                         return -ENOMEM;
2274         }
2275
2276         /* Check to ensure the stack will not grow into a hugetlb-only region */
2277         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2278                         vma->vm_end - size;
2279         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2280                 return -EFAULT;
2281
2282         /*
2283          * Overcommit..  This must be the final test, as it will
2284          * update security statistics.
2285          */
2286         if (security_vm_enough_memory_mm(mm, grow))
2287                 return -ENOMEM;
2288
2289         return 0;
2290 }
2291
2292 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2293 /*
2294  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2295  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2296  */
2297 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2298 {
2299         struct mm_struct *mm = vma->vm_mm;
2300         struct vm_area_struct *next;
2301         unsigned long gap_addr;
2302         int error = 0;
2303
2304         if (!(vma->vm_flags & VM_GROWSUP))
2305                 return -EFAULT;
2306
2307         /* Guard against exceeding limits of the address space. */
2308         address &= PAGE_MASK;
2309         if (address >= (TASK_SIZE & PAGE_MASK))
2310                 return -ENOMEM;
2311         address += PAGE_SIZE;
2312
2313         /* Enforce stack_guard_gap */
2314         gap_addr = address + stack_guard_gap;
2315
2316         /* Guard against overflow */
2317         if (gap_addr < address || gap_addr > TASK_SIZE)
2318                 gap_addr = TASK_SIZE;
2319
2320         next = vma->vm_next;
2321         if (next && next->vm_start < gap_addr &&
2322                         (next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2323                 if (!(next->vm_flags & VM_GROWSUP))
2324                         return -ENOMEM;
2325                 /* Check that both stack segments have the same anon_vma? */
2326         }
2327
2328         /* We must make sure the anon_vma is allocated. */
2329         if (unlikely(anon_vma_prepare(vma)))
2330                 return -ENOMEM;
2331
2332         /*
2333          * vma->vm_start/vm_end cannot change under us because the caller
2334          * is required to hold the mmap_sem in read mode.  We need the
2335          * anon_vma lock to serialize against concurrent expand_stacks.
2336          */
2337         anon_vma_lock_write(vma->anon_vma);
2338
2339         /* Somebody else might have raced and expanded it already */
2340         if (address > vma->vm_end) {
2341                 unsigned long size, grow;
2342
2343                 size = address - vma->vm_start;
2344                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2345
2346                 error = -ENOMEM;
2347                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2348                         error = acct_stack_growth(vma, size, grow);
2349                         if (!error) {
2350                                 /*
2351                                  * vma_gap_update() doesn't support concurrent
2352                                  * updates, but we only hold a shared mmap_sem
2353                                  * lock here, so we need to protect against
2354                                  * concurrent vma expansions.
2355                                  * anon_vma_lock_write() doesn't help here, as
2356                                  * we don't guarantee that all growable vmas
2357                                  * in a mm share the same root anon vma.
2358                                  * So, we reuse mm->page_table_lock to guard
2359                                  * against concurrent vma expansions.
2360                                  */
2361                                 spin_lock(&mm->page_table_lock);
2362                                 if (vma->vm_flags & VM_LOCKED)
2363                                         mm->locked_vm += grow;
2364                                 vm_stat_account(mm, vma->vm_flags, grow);
2365                                 anon_vma_interval_tree_pre_update_vma(vma);
2366                                 vma->vm_end = address;
2367                                 anon_vma_interval_tree_post_update_vma(vma);
2368                                 if (vma->vm_next)
2369                                         vma_gap_update(vma->vm_next);
2370                                 else
2371                                         mm->highest_vm_end = vm_end_gap(vma);
2372                                 spin_unlock(&mm->page_table_lock);
2373
2374                                 perf_event_mmap(vma);
2375                         }
2376                 }
2377         }
2378         anon_vma_unlock_write(vma->anon_vma);
2379         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2380         validate_mm(mm);
2381         return error;
2382 }
2383 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2384
2385 /*
2386  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2387  */
2388 int expand_downwards(struct vm_area_struct *vma,
2389                                    unsigned long address)
2390 {
2391         struct mm_struct *mm = vma->vm_mm;
2392         struct vm_area_struct *prev;
2393         int error;
2394
2395         address &= PAGE_MASK;
2396         error = security_mmap_addr(address);
2397         if (error)
2398                 return error;
2399
2400         /* Enforce stack_guard_gap */
2401         prev = vma->vm_prev;
2402         /* Check that both stack segments have the same anon_vma? */
2403         if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2404                         (prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2405                 if (address - prev->vm_end < stack_guard_gap)
2406                         return -ENOMEM;
2407         }
2408
2409         /* We must make sure the anon_vma is allocated. */
2410         if (unlikely(anon_vma_prepare(vma)))
2411                 return -ENOMEM;
2412
2413         /*
2414          * vma->vm_start/vm_end cannot change under us because the caller
2415          * is required to hold the mmap_sem in read mode.  We need the
2416          * anon_vma lock to serialize against concurrent expand_stacks.
2417          */
2418         anon_vma_lock_write(vma->anon_vma);
2419
2420         /* Somebody else might have raced and expanded it already */
2421         if (address < vma->vm_start) {
2422                 unsigned long size, grow;
2423
2424                 size = vma->vm_end - address;
2425                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2426
2427                 error = -ENOMEM;
2428                 if (grow <= vma->vm_pgoff) {
2429                         error = acct_stack_growth(vma, size, grow);
2430                         if (!error) {
2431                                 /*
2432                                  * vma_gap_update() doesn't support concurrent
2433                                  * updates, but we only hold a shared mmap_sem
2434                                  * lock here, so we need to protect against
2435                                  * concurrent vma expansions.
2436                                  * anon_vma_lock_write() doesn't help here, as
2437                                  * we don't guarantee that all growable vmas
2438                                  * in a mm share the same root anon vma.
2439                                  * So, we reuse mm->page_table_lock to guard
2440                                  * against concurrent vma expansions.
2441                                  */
2442                                 spin_lock(&mm->page_table_lock);
2443                                 if (vma->vm_flags & VM_LOCKED)
2444                                         mm->locked_vm += grow;
2445                                 vm_stat_account(mm, vma->vm_flags, grow);
2446                                 anon_vma_interval_tree_pre_update_vma(vma);
2447                                 vma->vm_start = address;
2448                                 vma->vm_pgoff -= grow;
2449                                 anon_vma_interval_tree_post_update_vma(vma);
2450                                 vma_gap_update(vma);
2451                                 spin_unlock(&mm->page_table_lock);
2452
2453                                 perf_event_mmap(vma);
2454                         }
2455                 }
2456         }
2457         anon_vma_unlock_write(vma->anon_vma);
2458         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2459         validate_mm(mm);
2460         return error;
2461 }
2462
2463 /* enforced gap between the expanding stack and other mappings. */
2464 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2465
2466 static int __init cmdline_parse_stack_guard_gap(char *p)
2467 {
2468         unsigned long val;
2469         char *endptr;
2470
2471         val = simple_strtoul(p, &endptr, 10);
2472         if (!*endptr)
2473                 stack_guard_gap = val << PAGE_SHIFT;
2474
2475         return 0;
2476 }
2477 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2478
2479 #ifdef CONFIG_STACK_GROWSUP
2480 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2481 {
2482         return expand_upwards(vma, address);
2483 }
2484
2485 struct vm_area_struct *
2486 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2487 {
2488         struct vm_area_struct *vma, *prev;
2489
2490         addr &= PAGE_MASK;
2491         vma = find_vma_prev(mm, addr, &prev);
2492         if (vma && (vma->vm_start <= addr))
2493                 return vma;
2494         if (!prev || expand_stack(prev, addr))
2495                 return NULL;
2496         if (prev->vm_flags & VM_LOCKED)
2497                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2498         return prev;
2499 }
2500 #else
2501 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2502 {
2503         return expand_downwards(vma, address);
2504 }
2505
2506 struct vm_area_struct *
2507 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2508 {
2509         struct vm_area_struct *vma;
2510         unsigned long start;
2511
2512         addr &= PAGE_MASK;
2513         vma = find_vma(mm, addr);
2514         if (!vma)
2515                 return NULL;
2516         if (vma->vm_start <= addr)
2517                 return vma;
2518         if (!(vma->vm_flags & VM_GROWSDOWN))
2519                 return NULL;
2520         start = vma->vm_start;
2521         if (expand_stack(vma, addr))
2522                 return NULL;
2523         if (vma->vm_flags & VM_LOCKED)
2524                 populate_vma_page_range(vma, addr, start, NULL);
2525         return vma;
2526 }
2527 #endif
2528
2529 EXPORT_SYMBOL_GPL(find_extend_vma);
2530
2531 /*
2532  * Ok - we have the memory areas we should free on the vma list,
2533  * so release them, and do the vma updates.
2534  *
2535  * Called with the mm semaphore held.
2536  */
2537 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2538 {
2539         unsigned long nr_accounted = 0;
2540
2541         /* Update high watermark before we lower total_vm */
2542         update_hiwater_vm(mm);
2543         do {
2544                 long nrpages = vma_pages(vma);
2545
2546                 if (vma->vm_flags & VM_ACCOUNT)
2547                         nr_accounted += nrpages;
2548                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2549                 vma = remove_vma(vma);
2550         } while (vma);
2551         vm_unacct_memory(nr_accounted);
2552         validate_mm(mm);
2553 }
2554
2555 /*
2556  * Get rid of page table information in the indicated region.
2557  *
2558  * Called with the mm semaphore held.
2559  */
2560 static void unmap_region(struct mm_struct *mm,
2561                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2562                 unsigned long start, unsigned long end)
2563 {
2564         struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2565         struct mmu_gather tlb;
2566
2567         lru_add_drain();
2568         tlb_gather_mmu(&tlb, mm, start, end);
2569         update_hiwater_rss(mm);
2570         unmap_vmas(&tlb, vma, start, end);
2571         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2572                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2573         tlb_finish_mmu(&tlb, start, end);
2574 }
2575
2576 /*
2577  * Create a list of vma's touched by the unmap, removing them from the mm's
2578  * vma list as we go..
2579  */
2580 static void
2581 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2582         struct vm_area_struct *prev, unsigned long end)
2583 {
2584         struct vm_area_struct **insertion_point;
2585         struct vm_area_struct *tail_vma = NULL;
2586
2587         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2588         vma->vm_prev = NULL;
2589         do {
2590                 vma_rb_erase(vma, &mm->mm_rb);
2591                 mm->map_count--;
2592                 tail_vma = vma;
2593                 vma = vma->vm_next;
2594         } while (vma && vma->vm_start < end);
2595         *insertion_point = vma;
2596         if (vma) {
2597                 vma->vm_prev = prev;
2598                 vma_gap_update(vma);
2599         } else
2600                 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2601         tail_vma->vm_next = NULL;
2602
2603         /* Kill the cache */
2604         vmacache_invalidate(mm);
2605 }
2606
2607 /*
2608  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2609  * has already been checked or doesn't make sense to fail.
2610  */
2611 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2612                 unsigned long addr, int new_below)
2613 {
2614         struct vm_area_struct *new;
2615         int err;
2616
2617         if (vma->vm_ops && vma->vm_ops->split) {
2618                 err = vma->vm_ops->split(vma, addr);
2619                 if (err)
2620                         return err;
2621         }
2622
2623         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2624         if (!new)
2625                 return -ENOMEM;
2626
2627         /* most fields are the same, copy all, and then fixup */
2628         *new = *vma;
2629
2630         INIT_LIST_HEAD(&new->anon_vma_chain);
2631
2632         if (new_below)
2633                 new->vm_end = addr;
2634         else {
2635                 new->vm_start = addr;
2636                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2637         }
2638
2639         err = vma_dup_policy(vma, new);
2640         if (err)
2641                 goto out_free_vma;
2642
2643         err = anon_vma_clone(new, vma);
2644         if (err)
2645                 goto out_free_mpol;
2646
2647         if (new->vm_file)
2648                 get_file(new->vm_file);
2649
2650         if (new->vm_ops && new->vm_ops->open)
2651                 new->vm_ops->open(new);
2652
2653         if (new_below)
2654                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2655                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2656         else
2657                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2658
2659         /* Success. */
2660         if (!err)
2661                 return 0;
2662
2663         /* Clean everything up if vma_adjust failed. */
2664         if (new->vm_ops && new->vm_ops->close)
2665                 new->vm_ops->close(new);
2666         if (new->vm_file)
2667                 fput(new->vm_file);
2668         unlink_anon_vmas(new);
2669  out_free_mpol:
2670         mpol_put(vma_policy(new));
2671  out_free_vma:
2672         kmem_cache_free(vm_area_cachep, new);
2673         return err;
2674 }
2675
2676 /*
2677  * Split a vma into two pieces at address 'addr', a new vma is allocated
2678  * either for the first part or the tail.
2679  */
2680 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2681               unsigned long addr, int new_below)
2682 {
2683         if (mm->map_count >= sysctl_max_map_count)
2684                 return -ENOMEM;
2685
2686         return __split_vma(mm, vma, addr, new_below);
2687 }
2688
2689 /* Munmap is split into 2 main parts -- this part which finds
2690  * what needs doing, and the areas themselves, which do the
2691  * work.  This now handles partial unmappings.
2692  * Jeremy Fitzhardinge <jeremy@goop.org>
2693  */
2694 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2695               struct list_head *uf)
2696 {
2697         unsigned long end;
2698         struct vm_area_struct *vma, *prev, *last;
2699
2700         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2701                 return -EINVAL;
2702
2703         len = PAGE_ALIGN(len);
2704         if (len == 0)
2705                 return -EINVAL;
2706
2707         /* Find the first overlapping VMA */
2708         vma = find_vma(mm, start);
2709         if (!vma)
2710                 return 0;
2711         prev = vma->vm_prev;
2712         /* we have  start < vma->vm_end  */
2713
2714         /* if it doesn't overlap, we have nothing.. */
2715         end = start + len;
2716         if (vma->vm_start >= end)
2717                 return 0;
2718
2719         /*
2720          * If we need to split any vma, do it now to save pain later.
2721          *
2722          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2723          * unmapped vm_area_struct will remain in use: so lower split_vma
2724          * places tmp vma above, and higher split_vma places tmp vma below.
2725          */
2726         if (start > vma->vm_start) {
2727                 int error;
2728
2729                 /*
2730                  * Make sure that map_count on return from munmap() will
2731                  * not exceed its limit; but let map_count go just above
2732                  * its limit temporarily, to help free resources as expected.
2733                  */
2734                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2735                         return -ENOMEM;
2736
2737                 error = __split_vma(mm, vma, start, 0);
2738                 if (error)
2739                         return error;
2740                 prev = vma;
2741         }
2742
2743         /* Does it split the last one? */
2744         last = find_vma(mm, end);
2745         if (last && end > last->vm_start) {
2746                 int error = __split_vma(mm, last, end, 1);
2747                 if (error)
2748                         return error;
2749         }
2750         vma = prev ? prev->vm_next : mm->mmap;
2751
2752         if (unlikely(uf)) {
2753                 /*
2754                  * If userfaultfd_unmap_prep returns an error the vmas
2755                  * will remain splitted, but userland will get a
2756                  * highly unexpected error anyway. This is no
2757                  * different than the case where the first of the two
2758                  * __split_vma fails, but we don't undo the first
2759                  * split, despite we could. This is unlikely enough
2760                  * failure that it's not worth optimizing it for.
2761                  */
2762                 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2763                 if (error)
2764                         return error;
2765         }
2766
2767         /*
2768          * unlock any mlock()ed ranges before detaching vmas
2769          */
2770         if (mm->locked_vm) {
2771                 struct vm_area_struct *tmp = vma;
2772                 while (tmp && tmp->vm_start < end) {
2773                         if (tmp->vm_flags & VM_LOCKED) {
2774                                 mm->locked_vm -= vma_pages(tmp);
2775                                 munlock_vma_pages_all(tmp);
2776                         }
2777                         tmp = tmp->vm_next;
2778                 }
2779         }
2780
2781         /*
2782          * Remove the vma's, and unmap the actual pages
2783          */
2784         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2785         unmap_region(mm, vma, prev, start, end);
2786
2787         arch_unmap(mm, vma, start, end);
2788
2789         /* Fix up all other VM information */
2790         remove_vma_list(mm, vma);
2791
2792         return 0;
2793 }
2794
2795 int vm_munmap(unsigned long start, size_t len)
2796 {
2797         int ret;
2798         struct mm_struct *mm = current->mm;
2799         LIST_HEAD(uf);
2800
2801         if (down_write_killable(&mm->mmap_sem))
2802                 return -EINTR;
2803
2804         ret = do_munmap(mm, start, len, &uf);
2805         up_write(&mm->mmap_sem);
2806         userfaultfd_unmap_complete(mm, &uf);
2807         return ret;
2808 }
2809 EXPORT_SYMBOL(vm_munmap);
2810
2811 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2812 {
2813         profile_munmap(addr);
2814         return vm_munmap(addr, len);
2815 }
2816
2817
2818 /*
2819  * Emulation of deprecated remap_file_pages() syscall.
2820  */
2821 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2822                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2823 {
2824
2825         struct mm_struct *mm = current->mm;
2826         struct vm_area_struct *vma;
2827         unsigned long populate = 0;
2828         unsigned long ret = -EINVAL;
2829         struct file *file;
2830
2831         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2832                      current->comm, current->pid);
2833
2834         if (prot)
2835                 return ret;
2836         start = start & PAGE_MASK;
2837         size = size & PAGE_MASK;
2838
2839         if (start + size <= start)
2840                 return ret;
2841
2842         /* Does pgoff wrap? */
2843         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2844                 return ret;
2845
2846         if (down_write_killable(&mm->mmap_sem))
2847                 return -EINTR;
2848
2849         vma = find_vma(mm, start);
2850
2851         if (!vma || !(vma->vm_flags & VM_SHARED))
2852                 goto out;
2853
2854         if (start < vma->vm_start)
2855                 goto out;
2856
2857         if (start + size > vma->vm_end) {
2858                 struct vm_area_struct *next;
2859
2860                 for (next = vma->vm_next; next; next = next->vm_next) {
2861                         /* hole between vmas ? */
2862                         if (next->vm_start != next->vm_prev->vm_end)
2863                                 goto out;
2864
2865                         if (next->vm_file != vma->vm_file)
2866                                 goto out;
2867
2868                         if (next->vm_flags != vma->vm_flags)
2869                                 goto out;
2870
2871                         if (start + size <= next->vm_end)
2872                                 break;
2873                 }
2874
2875                 if (!next)
2876                         goto out;
2877         }
2878
2879         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2880         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2881         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2882
2883         flags &= MAP_NONBLOCK;
2884         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2885         if (vma->vm_flags & VM_LOCKED) {
2886                 struct vm_area_struct *tmp;
2887                 flags |= MAP_LOCKED;
2888
2889                 /* drop PG_Mlocked flag for over-mapped range */
2890                 for (tmp = vma; tmp->vm_start >= start + size;
2891                                 tmp = tmp->vm_next) {
2892                         /*
2893                          * Split pmd and munlock page on the border
2894                          * of the range.
2895                          */
2896                         vma_adjust_trans_huge(tmp, start, start + size, 0);
2897
2898                         munlock_vma_pages_range(tmp,
2899                                         max(tmp->vm_start, start),
2900                                         min(tmp->vm_end, start + size));
2901                 }
2902         }
2903
2904         file = get_file(vma->vm_file);
2905         ret = do_mmap_pgoff(vma->vm_file, start, size,
2906                         prot, flags, pgoff, &populate, NULL);
2907         fput(file);
2908 out:
2909         up_write(&mm->mmap_sem);
2910         if (populate)
2911                 mm_populate(ret, populate);
2912         if (!IS_ERR_VALUE(ret))
2913                 ret = 0;
2914         return ret;
2915 }
2916
2917 static inline void verify_mm_writelocked(struct mm_struct *mm)
2918 {
2919 #ifdef CONFIG_DEBUG_VM
2920         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2921                 WARN_ON(1);
2922                 up_read(&mm->mmap_sem);
2923         }
2924 #endif
2925 }
2926
2927 /*
2928  *  this is really a simplified "do_mmap".  it only handles
2929  *  anonymous maps.  eventually we may be able to do some
2930  *  brk-specific accounting here.
2931  */
2932 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags, struct list_head *uf)
2933 {
2934         struct mm_struct *mm = current->mm;
2935         struct vm_area_struct *vma, *prev;
2936         unsigned long len;
2937         struct rb_node **rb_link, *rb_parent;
2938         pgoff_t pgoff = addr >> PAGE_SHIFT;
2939         int error;
2940
2941         len = PAGE_ALIGN(request);
2942         if (len < request)
2943                 return -ENOMEM;
2944         if (!len)
2945                 return 0;
2946
2947         /* Until we need other flags, refuse anything except VM_EXEC. */
2948         if ((flags & (~VM_EXEC)) != 0)
2949                 return -EINVAL;
2950         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2951
2952         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2953         if (offset_in_page(error))
2954                 return error;
2955
2956         error = mlock_future_check(mm, mm->def_flags, len);
2957         if (error)
2958                 return error;
2959
2960         /*
2961          * mm->mmap_sem is required to protect against another thread
2962          * changing the mappings in case we sleep.
2963          */
2964         verify_mm_writelocked(mm);
2965
2966         /*
2967          * Clear old maps.  this also does some error checking for us
2968          */
2969         while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2970                               &rb_parent)) {
2971                 if (do_munmap(mm, addr, len, uf))
2972                         return -ENOMEM;
2973         }
2974
2975         /* Check against address space limits *after* clearing old maps... */
2976         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2977                 return -ENOMEM;
2978
2979         if (mm->map_count > sysctl_max_map_count)
2980                 return -ENOMEM;
2981
2982         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2983                 return -ENOMEM;
2984
2985         /* Can we just expand an old private anonymous mapping? */
2986         vma = vma_merge(mm, prev, addr, addr + len, flags,
2987                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2988         if (vma)
2989                 goto out;
2990
2991         /*
2992          * create a vma struct for an anonymous mapping
2993          */
2994         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2995         if (!vma) {
2996                 vm_unacct_memory(len >> PAGE_SHIFT);
2997                 return -ENOMEM;
2998         }
2999
3000         INIT_LIST_HEAD(&vma->anon_vma_chain);
3001         vma->vm_mm = mm;
3002         vma->vm_start = addr;
3003         vma->vm_end = addr + len;
3004         vma->vm_pgoff = pgoff;
3005         vma->vm_flags = flags;
3006         vma->vm_page_prot = vm_get_page_prot(flags);
3007         vma_link(mm, vma, prev, rb_link, rb_parent);
3008 out:
3009         perf_event_mmap(vma);
3010         mm->total_vm += len >> PAGE_SHIFT;
3011         mm->data_vm += len >> PAGE_SHIFT;
3012         if (flags & VM_LOCKED)
3013                 mm->locked_vm += (len >> PAGE_SHIFT);
3014         vma->vm_flags |= VM_SOFTDIRTY;
3015         return 0;
3016 }
3017
3018 static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf)
3019 {
3020         return do_brk_flags(addr, len, 0, uf);
3021 }
3022
3023 int vm_brk_flags(unsigned long addr, unsigned long len, unsigned long flags)
3024 {
3025         struct mm_struct *mm = current->mm;
3026         int ret;
3027         bool populate;
3028         LIST_HEAD(uf);
3029
3030         if (down_write_killable(&mm->mmap_sem))
3031                 return -EINTR;
3032
3033         ret = do_brk_flags(addr, len, flags, &uf);
3034         populate = ((mm->def_flags & VM_LOCKED) != 0);
3035         up_write(&mm->mmap_sem);
3036         userfaultfd_unmap_complete(mm, &uf);
3037         if (populate && !ret)
3038                 mm_populate(addr, len);
3039         return ret;
3040 }
3041 EXPORT_SYMBOL(vm_brk_flags);
3042
3043 int vm_brk(unsigned long addr, unsigned long len)
3044 {
3045         return vm_brk_flags(addr, len, 0);
3046 }
3047 EXPORT_SYMBOL(vm_brk);
3048
3049 /* Release all mmaps. */
3050 void exit_mmap(struct mm_struct *mm)
3051 {
3052         struct mmu_gather tlb;
3053         struct vm_area_struct *vma;
3054         unsigned long nr_accounted = 0;
3055
3056         /* mm's last user has gone, and its about to be pulled down */
3057         mmu_notifier_release(mm);
3058
3059         if (unlikely(mm_is_oom_victim(mm))) {
3060                 /*
3061                  * Manually reap the mm to free as much memory as possible.
3062                  * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3063                  * this mm from further consideration.  Taking mm->mmap_sem for
3064                  * write after setting MMF_OOM_SKIP will guarantee that the oom
3065                  * reaper will not run on this mm again after mmap_sem is
3066                  * dropped.
3067                  *
3068                  * Nothing can be holding mm->mmap_sem here and the above call
3069                  * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3070                  * __oom_reap_task_mm() will not block.
3071                  *
3072                  * This needs to be done before calling munlock_vma_pages_all(),
3073                  * which clears VM_LOCKED, otherwise the oom reaper cannot
3074                  * reliably test it.
3075                  */
3076                 mutex_lock(&oom_lock);
3077                 __oom_reap_task_mm(mm);
3078                 mutex_unlock(&oom_lock);
3079
3080                 set_bit(MMF_OOM_SKIP, &mm->flags);
3081                 down_write(&mm->mmap_sem);
3082                 up_write(&mm->mmap_sem);
3083         }
3084
3085         if (mm->locked_vm) {
3086                 vma = mm->mmap;
3087                 while (vma) {
3088                         if (vma->vm_flags & VM_LOCKED)
3089                                 munlock_vma_pages_all(vma);
3090                         vma = vma->vm_next;
3091                 }
3092         }
3093
3094         arch_exit_mmap(mm);
3095
3096         vma = mm->mmap;
3097         if (!vma)       /* Can happen if dup_mmap() received an OOM */
3098                 return;
3099
3100         lru_add_drain();
3101         flush_cache_mm(mm);
3102         tlb_gather_mmu(&tlb, mm, 0, -1);
3103         /* update_hiwater_rss(mm) here? but nobody should be looking */
3104         /* Use -1 here to ensure all VMAs in the mm are unmapped */
3105         unmap_vmas(&tlb, vma, 0, -1);
3106         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3107         tlb_finish_mmu(&tlb, 0, -1);
3108
3109         /*
3110          * Walk the list again, actually closing and freeing it,
3111          * with preemption enabled, without holding any MM locks.
3112          */
3113         while (vma) {
3114                 if (vma->vm_flags & VM_ACCOUNT)
3115                         nr_accounted += vma_pages(vma);
3116                 vma = remove_vma(vma);
3117         }
3118         vm_unacct_memory(nr_accounted);
3119 }
3120
3121 /* Insert vm structure into process list sorted by address
3122  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3123  * then i_mmap_rwsem is taken here.
3124  */
3125 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3126 {
3127         struct vm_area_struct *prev;
3128         struct rb_node **rb_link, *rb_parent;
3129
3130         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3131                            &prev, &rb_link, &rb_parent))
3132                 return -ENOMEM;
3133         if ((vma->vm_flags & VM_ACCOUNT) &&
3134              security_vm_enough_memory_mm(mm, vma_pages(vma)))
3135                 return -ENOMEM;
3136
3137         /*
3138          * The vm_pgoff of a purely anonymous vma should be irrelevant
3139          * until its first write fault, when page's anon_vma and index
3140          * are set.  But now set the vm_pgoff it will almost certainly
3141          * end up with (unless mremap moves it elsewhere before that
3142          * first wfault), so /proc/pid/maps tells a consistent story.
3143          *
3144          * By setting it to reflect the virtual start address of the
3145          * vma, merges and splits can happen in a seamless way, just
3146          * using the existing file pgoff checks and manipulations.
3147          * Similarly in do_mmap_pgoff and in do_brk.
3148          */
3149         if (vma_is_anonymous(vma)) {
3150                 BUG_ON(vma->anon_vma);
3151                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3152         }
3153
3154         vma_link(mm, vma, prev, rb_link, rb_parent);
3155         return 0;
3156 }
3157
3158 /*
3159  * Copy the vma structure to a new location in the same mm,
3160  * prior to moving page table entries, to effect an mremap move.
3161  */
3162 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3163         unsigned long addr, unsigned long len, pgoff_t pgoff,
3164         bool *need_rmap_locks)
3165 {
3166         struct vm_area_struct *vma = *vmap;
3167         unsigned long vma_start = vma->vm_start;
3168         struct mm_struct *mm = vma->vm_mm;
3169         struct vm_area_struct *new_vma, *prev;
3170         struct rb_node **rb_link, *rb_parent;
3171         bool faulted_in_anon_vma = true;
3172
3173         /*
3174          * If anonymous vma has not yet been faulted, update new pgoff
3175          * to match new location, to increase its chance of merging.
3176          */
3177         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3178                 pgoff = addr >> PAGE_SHIFT;
3179                 faulted_in_anon_vma = false;
3180         }
3181
3182         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3183                 return NULL;    /* should never get here */
3184         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3185                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3186                             vma->vm_userfaultfd_ctx);
3187         if (new_vma) {
3188                 /*
3189                  * Source vma may have been merged into new_vma
3190                  */
3191                 if (unlikely(vma_start >= new_vma->vm_start &&
3192                              vma_start < new_vma->vm_end)) {
3193                         /*
3194                          * The only way we can get a vma_merge with
3195                          * self during an mremap is if the vma hasn't
3196                          * been faulted in yet and we were allowed to
3197                          * reset the dst vma->vm_pgoff to the
3198                          * destination address of the mremap to allow
3199                          * the merge to happen. mremap must change the
3200                          * vm_pgoff linearity between src and dst vmas
3201                          * (in turn preventing a vma_merge) to be
3202                          * safe. It is only safe to keep the vm_pgoff
3203                          * linear if there are no pages mapped yet.
3204                          */
3205                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3206                         *vmap = vma = new_vma;
3207                 }
3208                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3209         } else {
3210                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3211                 if (!new_vma)
3212                         goto out;
3213                 *new_vma = *vma;
3214                 new_vma->vm_start = addr;
3215                 new_vma->vm_end = addr + len;
3216                 new_vma->vm_pgoff = pgoff;
3217                 if (vma_dup_policy(vma, new_vma))
3218                         goto out_free_vma;
3219                 INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3220                 if (anon_vma_clone(new_vma, vma))
3221                         goto out_free_mempol;
3222                 if (new_vma->vm_file)
3223                         get_file(new_vma->vm_file);
3224                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3225                         new_vma->vm_ops->open(new_vma);
3226                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3227                 *need_rmap_locks = false;
3228         }
3229         return new_vma;
3230
3231 out_free_mempol:
3232         mpol_put(vma_policy(new_vma));
3233 out_free_vma:
3234         kmem_cache_free(vm_area_cachep, new_vma);
3235 out:
3236         return NULL;
3237 }
3238
3239 /*
3240  * Return true if the calling process may expand its vm space by the passed
3241  * number of pages
3242  */
3243 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3244 {
3245         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3246                 return false;
3247
3248         if (is_data_mapping(flags) &&
3249             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3250                 /* Workaround for Valgrind */
3251                 if (rlimit(RLIMIT_DATA) == 0 &&
3252                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3253                         return true;
3254
3255                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3256                              current->comm, current->pid,
3257                              (mm->data_vm + npages) << PAGE_SHIFT,
3258                              rlimit(RLIMIT_DATA),
3259                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3260
3261                 if (!ignore_rlimit_data)
3262                         return false;
3263         }
3264
3265         return true;
3266 }
3267
3268 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3269 {
3270         mm->total_vm += npages;
3271
3272         if (is_exec_mapping(flags))
3273                 mm->exec_vm += npages;
3274         else if (is_stack_mapping(flags))
3275                 mm->stack_vm += npages;
3276         else if (is_data_mapping(flags))
3277                 mm->data_vm += npages;
3278 }
3279
3280 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3281
3282 /*
3283  * Having a close hook prevents vma merging regardless of flags.
3284  */
3285 static void special_mapping_close(struct vm_area_struct *vma)
3286 {
3287 }
3288
3289 static const char *special_mapping_name(struct vm_area_struct *vma)
3290 {
3291         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3292 }
3293
3294 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3295 {
3296         struct vm_special_mapping *sm = new_vma->vm_private_data;
3297
3298         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3299                 return -EFAULT;
3300
3301         if (sm->mremap)
3302                 return sm->mremap(sm, new_vma);
3303
3304         return 0;
3305 }
3306
3307 static const struct vm_operations_struct special_mapping_vmops = {
3308         .close = special_mapping_close,
3309         .fault = special_mapping_fault,
3310         .mremap = special_mapping_mremap,
3311         .name = special_mapping_name,
3312 };
3313
3314 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3315         .close = special_mapping_close,
3316         .fault = special_mapping_fault,
3317 };
3318
3319 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3320 {
3321         struct vm_area_struct *vma = vmf->vma;
3322         pgoff_t pgoff;
3323         struct page **pages;
3324
3325         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3326                 pages = vma->vm_private_data;
3327         } else {
3328                 struct vm_special_mapping *sm = vma->vm_private_data;
3329
3330                 if (sm->fault)
3331                         return sm->fault(sm, vmf->vma, vmf);
3332
3333                 pages = sm->pages;
3334         }
3335
3336         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3337                 pgoff--;
3338
3339         if (*pages) {
3340                 struct page *page = *pages;
3341                 get_page(page);
3342                 vmf->page = page;
3343                 return 0;
3344         }
3345
3346         return VM_FAULT_SIGBUS;
3347 }
3348
3349 static struct vm_area_struct *__install_special_mapping(
3350         struct mm_struct *mm,
3351         unsigned long addr, unsigned long len,
3352         unsigned long vm_flags, void *priv,
3353         const struct vm_operations_struct *ops)
3354 {
3355         int ret;
3356         struct vm_area_struct *vma;
3357
3358         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3359         if (unlikely(vma == NULL))
3360                 return ERR_PTR(-ENOMEM);
3361
3362         INIT_LIST_HEAD(&vma->anon_vma_chain);
3363         vma->vm_mm = mm;
3364         vma->vm_start = addr;
3365         vma->vm_end = addr + len;
3366
3367         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3368         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3369
3370         vma->vm_ops = ops;
3371         vma->vm_private_data = priv;
3372
3373         ret = insert_vm_struct(mm, vma);
3374         if (ret)
3375                 goto out;
3376
3377         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3378
3379         perf_event_mmap(vma);
3380
3381         return vma;
3382
3383 out:
3384         kmem_cache_free(vm_area_cachep, vma);
3385         return ERR_PTR(ret);
3386 }
3387
3388 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3389         const struct vm_special_mapping *sm)
3390 {
3391         return vma->vm_private_data == sm &&
3392                 (vma->vm_ops == &special_mapping_vmops ||
3393                  vma->vm_ops == &legacy_special_mapping_vmops);
3394 }
3395
3396 /*
3397  * Called with mm->mmap_sem held for writing.
3398  * Insert a new vma covering the given region, with the given flags.
3399  * Its pages are supplied by the given array of struct page *.
3400  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3401  * The region past the last page supplied will always produce SIGBUS.
3402  * The array pointer and the pages it points to are assumed to stay alive
3403  * for as long as this mapping might exist.
3404  */
3405 struct vm_area_struct *_install_special_mapping(
3406         struct mm_struct *mm,
3407         unsigned long addr, unsigned long len,
3408         unsigned long vm_flags, const struct vm_special_mapping *spec)
3409 {
3410         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3411                                         &special_mapping_vmops);
3412 }
3413
3414 int install_special_mapping(struct mm_struct *mm,
3415                             unsigned long addr, unsigned long len,
3416                             unsigned long vm_flags, struct page **pages)
3417 {
3418         struct vm_area_struct *vma = __install_special_mapping(
3419                 mm, addr, len, vm_flags, (void *)pages,
3420                 &legacy_special_mapping_vmops);
3421
3422         return PTR_ERR_OR_ZERO(vma);
3423 }
3424
3425 static DEFINE_MUTEX(mm_all_locks_mutex);
3426
3427 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3428 {
3429         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3430                 /*
3431                  * The LSB of head.next can't change from under us
3432                  * because we hold the mm_all_locks_mutex.
3433                  */
3434                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3435                 /*
3436                  * We can safely modify head.next after taking the
3437                  * anon_vma->root->rwsem. If some other vma in this mm shares
3438                  * the same anon_vma we won't take it again.
3439                  *
3440                  * No need of atomic instructions here, head.next
3441                  * can't change from under us thanks to the
3442                  * anon_vma->root->rwsem.
3443                  */
3444                 if (__test_and_set_bit(0, (unsigned long *)
3445                                        &anon_vma->root->rb_root.rb_root.rb_node))
3446                         BUG();
3447         }
3448 }
3449
3450 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3451 {
3452         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3453                 /*
3454                  * AS_MM_ALL_LOCKS can't change from under us because
3455                  * we hold the mm_all_locks_mutex.
3456                  *
3457                  * Operations on ->flags have to be atomic because
3458                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3459                  * mm_all_locks_mutex, there may be other cpus
3460                  * changing other bitflags in parallel to us.
3461                  */
3462                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3463                         BUG();
3464                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3465         }
3466 }
3467
3468 /*
3469  * This operation locks against the VM for all pte/vma/mm related
3470  * operations that could ever happen on a certain mm. This includes
3471  * vmtruncate, try_to_unmap, and all page faults.
3472  *
3473  * The caller must take the mmap_sem in write mode before calling
3474  * mm_take_all_locks(). The caller isn't allowed to release the
3475  * mmap_sem until mm_drop_all_locks() returns.
3476  *
3477  * mmap_sem in write mode is required in order to block all operations
3478  * that could modify pagetables and free pages without need of
3479  * altering the vma layout. It's also needed in write mode to avoid new
3480  * anon_vmas to be associated with existing vmas.
3481  *
3482  * A single task can't take more than one mm_take_all_locks() in a row
3483  * or it would deadlock.
3484  *
3485  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3486  * mapping->flags avoid to take the same lock twice, if more than one
3487  * vma in this mm is backed by the same anon_vma or address_space.
3488  *
3489  * We take locks in following order, accordingly to comment at beginning
3490  * of mm/rmap.c:
3491  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3492  *     hugetlb mapping);
3493  *   - all i_mmap_rwsem locks;
3494  *   - all anon_vma->rwseml
3495  *
3496  * We can take all locks within these types randomly because the VM code
3497  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3498  * mm_all_locks_mutex.
3499  *
3500  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3501  * that may have to take thousand of locks.
3502  *
3503  * mm_take_all_locks() can fail if it's interrupted by signals.
3504  */
3505 int mm_take_all_locks(struct mm_struct *mm)
3506 {
3507         struct vm_area_struct *vma;
3508         struct anon_vma_chain *avc;
3509
3510         BUG_ON(down_read_trylock(&mm->mmap_sem));
3511
3512         mutex_lock(&mm_all_locks_mutex);
3513
3514         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3515                 if (signal_pending(current))
3516                         goto out_unlock;
3517                 if (vma->vm_file && vma->vm_file->f_mapping &&
3518                                 is_vm_hugetlb_page(vma))
3519                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3520         }
3521
3522         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3523                 if (signal_pending(current))
3524                         goto out_unlock;
3525                 if (vma->vm_file && vma->vm_file->f_mapping &&
3526                                 !is_vm_hugetlb_page(vma))
3527                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3528         }
3529
3530         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3531                 if (signal_pending(current))
3532                         goto out_unlock;
3533                 if (vma->anon_vma)
3534                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3535                                 vm_lock_anon_vma(mm, avc->anon_vma);
3536         }
3537
3538         return 0;
3539
3540 out_unlock:
3541         mm_drop_all_locks(mm);
3542         return -EINTR;
3543 }
3544
3545 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3546 {
3547         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3548                 /*
3549                  * The LSB of head.next can't change to 0 from under
3550                  * us because we hold the mm_all_locks_mutex.
3551                  *
3552                  * We must however clear the bitflag before unlocking
3553                  * the vma so the users using the anon_vma->rb_root will
3554                  * never see our bitflag.
3555                  *
3556                  * No need of atomic instructions here, head.next
3557                  * can't change from under us until we release the
3558                  * anon_vma->root->rwsem.
3559                  */
3560                 if (!__test_and_clear_bit(0, (unsigned long *)
3561                                           &anon_vma->root->rb_root.rb_root.rb_node))
3562                         BUG();
3563                 anon_vma_unlock_write(anon_vma);
3564         }
3565 }
3566
3567 static void vm_unlock_mapping(struct address_space *mapping)
3568 {
3569         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3570                 /*
3571                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3572                  * because we hold the mm_all_locks_mutex.
3573                  */
3574                 i_mmap_unlock_write(mapping);
3575                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3576                                         &mapping->flags))
3577                         BUG();
3578         }
3579 }
3580
3581 /*
3582  * The mmap_sem cannot be released by the caller until
3583  * mm_drop_all_locks() returns.
3584  */
3585 void mm_drop_all_locks(struct mm_struct *mm)
3586 {
3587         struct vm_area_struct *vma;
3588         struct anon_vma_chain *avc;
3589
3590         BUG_ON(down_read_trylock(&mm->mmap_sem));
3591         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3592
3593         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3594                 if (vma->anon_vma)
3595                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3596                                 vm_unlock_anon_vma(avc->anon_vma);
3597                 if (vma->vm_file && vma->vm_file->f_mapping)
3598                         vm_unlock_mapping(vma->vm_file->f_mapping);
3599         }
3600
3601         mutex_unlock(&mm_all_locks_mutex);
3602 }
3603
3604 /*
3605  * initialise the percpu counter for VM
3606  */
3607 void __init mmap_init(void)
3608 {
3609         int ret;
3610
3611         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3612         VM_BUG_ON(ret);
3613 }
3614
3615 /*
3616  * Initialise sysctl_user_reserve_kbytes.
3617  *
3618  * This is intended to prevent a user from starting a single memory hogging
3619  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3620  * mode.
3621  *
3622  * The default value is min(3% of free memory, 128MB)
3623  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3624  */
3625 static int init_user_reserve(void)
3626 {
3627         unsigned long free_kbytes;
3628
3629         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3630
3631         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3632         return 0;
3633 }
3634 subsys_initcall(init_user_reserve);
3635
3636 /*
3637  * Initialise sysctl_admin_reserve_kbytes.
3638  *
3639  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3640  * to log in and kill a memory hogging process.
3641  *
3642  * Systems with more than 256MB will reserve 8MB, enough to recover
3643  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3644  * only reserve 3% of free pages by default.
3645  */
3646 static int init_admin_reserve(void)
3647 {
3648         unsigned long free_kbytes;
3649
3650         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3651
3652         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3653         return 0;
3654 }
3655 subsys_initcall(init_admin_reserve);
3656
3657 /*
3658  * Reinititalise user and admin reserves if memory is added or removed.
3659  *
3660  * The default user reserve max is 128MB, and the default max for the
3661  * admin reserve is 8MB. These are usually, but not always, enough to
3662  * enable recovery from a memory hogging process using login/sshd, a shell,
3663  * and tools like top. It may make sense to increase or even disable the
3664  * reserve depending on the existence of swap or variations in the recovery
3665  * tools. So, the admin may have changed them.
3666  *
3667  * If memory is added and the reserves have been eliminated or increased above
3668  * the default max, then we'll trust the admin.
3669  *
3670  * If memory is removed and there isn't enough free memory, then we
3671  * need to reset the reserves.
3672  *
3673  * Otherwise keep the reserve set by the admin.
3674  */
3675 static int reserve_mem_notifier(struct notifier_block *nb,
3676                              unsigned long action, void *data)
3677 {
3678         unsigned long tmp, free_kbytes;
3679
3680         switch (action) {
3681         case MEM_ONLINE:
3682                 /* Default max is 128MB. Leave alone if modified by operator. */
3683                 tmp = sysctl_user_reserve_kbytes;
3684                 if (0 < tmp && tmp < (1UL << 17))
3685                         init_user_reserve();
3686
3687                 /* Default max is 8MB.  Leave alone if modified by operator. */
3688                 tmp = sysctl_admin_reserve_kbytes;
3689                 if (0 < tmp && tmp < (1UL << 13))
3690                         init_admin_reserve();
3691
3692                 break;
3693         case MEM_OFFLINE:
3694                 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3695
3696                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3697                         init_user_reserve();
3698                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3699                                 sysctl_user_reserve_kbytes);
3700                 }
3701
3702                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3703                         init_admin_reserve();
3704                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3705                                 sysctl_admin_reserve_kbytes);
3706                 }
3707                 break;
3708         default:
3709                 break;
3710         }
3711         return NOTIFY_OK;
3712 }
3713
3714 static struct notifier_block reserve_mem_nb = {
3715         .notifier_call = reserve_mem_notifier,
3716 };
3717
3718 static int __meminit init_reserve_notifier(void)
3719 {
3720         if (register_hotmemory_notifier(&reserve_mem_nb))
3721                 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3722
3723         return 0;
3724 }
3725 subsys_initcall(init_reserve_notifier);