mm: memcg: make stats flushing threshold per-memcg
[linux-2.6-microblaze.git] / mm / mmap.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * mm/mmap.c
4  *
5  * Written by obz.
6  *
7  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
8  */
9
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/backing-dev.h>
15 #include <linux/mm.h>
16 #include <linux/mm_inline.h>
17 #include <linux/shm.h>
18 #include <linux/mman.h>
19 #include <linux/pagemap.h>
20 #include <linux/swap.h>
21 #include <linux/syscalls.h>
22 #include <linux/capability.h>
23 #include <linux/init.h>
24 #include <linux/file.h>
25 #include <linux/fs.h>
26 #include <linux/personality.h>
27 #include <linux/security.h>
28 #include <linux/hugetlb.h>
29 #include <linux/shmem_fs.h>
30 #include <linux/profile.h>
31 #include <linux/export.h>
32 #include <linux/mount.h>
33 #include <linux/mempolicy.h>
34 #include <linux/rmap.h>
35 #include <linux/mmu_notifier.h>
36 #include <linux/mmdebug.h>
37 #include <linux/perf_event.h>
38 #include <linux/audit.h>
39 #include <linux/khugepaged.h>
40 #include <linux/uprobes.h>
41 #include <linux/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 #include <linux/sched/mm.h>
49 #include <linux/ksm.h>
50
51 #include <linux/uaccess.h>
52 #include <asm/cacheflush.h>
53 #include <asm/tlb.h>
54 #include <asm/mmu_context.h>
55
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/mmap.h>
58
59 #include "internal.h"
60
61 #ifndef arch_mmap_check
62 #define arch_mmap_check(addr, len, flags)       (0)
63 #endif
64
65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
66 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
67 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
69 #endif
70 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
74 #endif
75
76 static bool ignore_rlimit_data;
77 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
78
79 static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
80                 struct vm_area_struct *vma, struct vm_area_struct *prev,
81                 struct vm_area_struct *next, unsigned long start,
82                 unsigned long end, unsigned long tree_end, bool mm_wr_locked);
83
84 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
85 {
86         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
87 }
88
89 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
90 void vma_set_page_prot(struct vm_area_struct *vma)
91 {
92         unsigned long vm_flags = vma->vm_flags;
93         pgprot_t vm_page_prot;
94
95         vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
96         if (vma_wants_writenotify(vma, vm_page_prot)) {
97                 vm_flags &= ~VM_SHARED;
98                 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
99         }
100         /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
101         WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
102 }
103
104 /*
105  * Requires inode->i_mapping->i_mmap_rwsem
106  */
107 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
108                 struct file *file, struct address_space *mapping)
109 {
110         if (vma_is_shared_maywrite(vma))
111                 mapping_unmap_writable(mapping);
112
113         flush_dcache_mmap_lock(mapping);
114         vma_interval_tree_remove(vma, &mapping->i_mmap);
115         flush_dcache_mmap_unlock(mapping);
116 }
117
118 /*
119  * Unlink a file-based vm structure from its interval tree, to hide
120  * vma from rmap and vmtruncate before freeing its page tables.
121  */
122 void unlink_file_vma(struct vm_area_struct *vma)
123 {
124         struct file *file = vma->vm_file;
125
126         if (file) {
127                 struct address_space *mapping = file->f_mapping;
128                 i_mmap_lock_write(mapping);
129                 __remove_shared_vm_struct(vma, file, mapping);
130                 i_mmap_unlock_write(mapping);
131         }
132 }
133
134 /*
135  * Close a vm structure and free it.
136  */
137 static void remove_vma(struct vm_area_struct *vma, bool unreachable)
138 {
139         might_sleep();
140         if (vma->vm_ops && vma->vm_ops->close)
141                 vma->vm_ops->close(vma);
142         if (vma->vm_file)
143                 fput(vma->vm_file);
144         mpol_put(vma_policy(vma));
145         if (unreachable)
146                 __vm_area_free(vma);
147         else
148                 vm_area_free(vma);
149 }
150
151 static inline struct vm_area_struct *vma_prev_limit(struct vma_iterator *vmi,
152                                                     unsigned long min)
153 {
154         return mas_prev(&vmi->mas, min);
155 }
156
157 /*
158  * check_brk_limits() - Use platform specific check of range & verify mlock
159  * limits.
160  * @addr: The address to check
161  * @len: The size of increase.
162  *
163  * Return: 0 on success.
164  */
165 static int check_brk_limits(unsigned long addr, unsigned long len)
166 {
167         unsigned long mapped_addr;
168
169         mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
170         if (IS_ERR_VALUE(mapped_addr))
171                 return mapped_addr;
172
173         return mlock_future_ok(current->mm, current->mm->def_flags, len)
174                 ? 0 : -EAGAIN;
175 }
176 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
177                 unsigned long addr, unsigned long request, unsigned long flags);
178 SYSCALL_DEFINE1(brk, unsigned long, brk)
179 {
180         unsigned long newbrk, oldbrk, origbrk;
181         struct mm_struct *mm = current->mm;
182         struct vm_area_struct *brkvma, *next = NULL;
183         unsigned long min_brk;
184         bool populate = false;
185         LIST_HEAD(uf);
186         struct vma_iterator vmi;
187
188         if (mmap_write_lock_killable(mm))
189                 return -EINTR;
190
191         origbrk = mm->brk;
192
193 #ifdef CONFIG_COMPAT_BRK
194         /*
195          * CONFIG_COMPAT_BRK can still be overridden by setting
196          * randomize_va_space to 2, which will still cause mm->start_brk
197          * to be arbitrarily shifted
198          */
199         if (current->brk_randomized)
200                 min_brk = mm->start_brk;
201         else
202                 min_brk = mm->end_data;
203 #else
204         min_brk = mm->start_brk;
205 #endif
206         if (brk < min_brk)
207                 goto out;
208
209         /*
210          * Check against rlimit here. If this check is done later after the test
211          * of oldbrk with newbrk then it can escape the test and let the data
212          * segment grow beyond its set limit the in case where the limit is
213          * not page aligned -Ram Gupta
214          */
215         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
216                               mm->end_data, mm->start_data))
217                 goto out;
218
219         newbrk = PAGE_ALIGN(brk);
220         oldbrk = PAGE_ALIGN(mm->brk);
221         if (oldbrk == newbrk) {
222                 mm->brk = brk;
223                 goto success;
224         }
225
226         /* Always allow shrinking brk. */
227         if (brk <= mm->brk) {
228                 /* Search one past newbrk */
229                 vma_iter_init(&vmi, mm, newbrk);
230                 brkvma = vma_find(&vmi, oldbrk);
231                 if (!brkvma || brkvma->vm_start >= oldbrk)
232                         goto out; /* mapping intersects with an existing non-brk vma. */
233                 /*
234                  * mm->brk must be protected by write mmap_lock.
235                  * do_vma_munmap() will drop the lock on success,  so update it
236                  * before calling do_vma_munmap().
237                  */
238                 mm->brk = brk;
239                 if (do_vma_munmap(&vmi, brkvma, newbrk, oldbrk, &uf, true))
240                         goto out;
241
242                 goto success_unlocked;
243         }
244
245         if (check_brk_limits(oldbrk, newbrk - oldbrk))
246                 goto out;
247
248         /*
249          * Only check if the next VMA is within the stack_guard_gap of the
250          * expansion area
251          */
252         vma_iter_init(&vmi, mm, oldbrk);
253         next = vma_find(&vmi, newbrk + PAGE_SIZE + stack_guard_gap);
254         if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
255                 goto out;
256
257         brkvma = vma_prev_limit(&vmi, mm->start_brk);
258         /* Ok, looks good - let it rip. */
259         if (do_brk_flags(&vmi, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
260                 goto out;
261
262         mm->brk = brk;
263         if (mm->def_flags & VM_LOCKED)
264                 populate = true;
265
266 success:
267         mmap_write_unlock(mm);
268 success_unlocked:
269         userfaultfd_unmap_complete(mm, &uf);
270         if (populate)
271                 mm_populate(oldbrk, newbrk - oldbrk);
272         return brk;
273
274 out:
275         mm->brk = origbrk;
276         mmap_write_unlock(mm);
277         return origbrk;
278 }
279
280 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
281 static void validate_mm(struct mm_struct *mm)
282 {
283         int bug = 0;
284         int i = 0;
285         struct vm_area_struct *vma;
286         VMA_ITERATOR(vmi, mm, 0);
287
288         mt_validate(&mm->mm_mt);
289         for_each_vma(vmi, vma) {
290 #ifdef CONFIG_DEBUG_VM_RB
291                 struct anon_vma *anon_vma = vma->anon_vma;
292                 struct anon_vma_chain *avc;
293 #endif
294                 unsigned long vmi_start, vmi_end;
295                 bool warn = 0;
296
297                 vmi_start = vma_iter_addr(&vmi);
298                 vmi_end = vma_iter_end(&vmi);
299                 if (VM_WARN_ON_ONCE_MM(vma->vm_end != vmi_end, mm))
300                         warn = 1;
301
302                 if (VM_WARN_ON_ONCE_MM(vma->vm_start != vmi_start, mm))
303                         warn = 1;
304
305                 if (warn) {
306                         pr_emerg("issue in %s\n", current->comm);
307                         dump_stack();
308                         dump_vma(vma);
309                         pr_emerg("tree range: %px start %lx end %lx\n", vma,
310                                  vmi_start, vmi_end - 1);
311                         vma_iter_dump_tree(&vmi);
312                 }
313
314 #ifdef CONFIG_DEBUG_VM_RB
315                 if (anon_vma) {
316                         anon_vma_lock_read(anon_vma);
317                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
318                                 anon_vma_interval_tree_verify(avc);
319                         anon_vma_unlock_read(anon_vma);
320                 }
321 #endif
322                 i++;
323         }
324         if (i != mm->map_count) {
325                 pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);
326                 bug = 1;
327         }
328         VM_BUG_ON_MM(bug, mm);
329 }
330
331 #else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
332 #define validate_mm(mm) do { } while (0)
333 #endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
334
335 /*
336  * vma has some anon_vma assigned, and is already inserted on that
337  * anon_vma's interval trees.
338  *
339  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
340  * vma must be removed from the anon_vma's interval trees using
341  * anon_vma_interval_tree_pre_update_vma().
342  *
343  * After the update, the vma will be reinserted using
344  * anon_vma_interval_tree_post_update_vma().
345  *
346  * The entire update must be protected by exclusive mmap_lock and by
347  * the root anon_vma's mutex.
348  */
349 static inline void
350 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
351 {
352         struct anon_vma_chain *avc;
353
354         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
355                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
356 }
357
358 static inline void
359 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
360 {
361         struct anon_vma_chain *avc;
362
363         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
364                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
365 }
366
367 static unsigned long count_vma_pages_range(struct mm_struct *mm,
368                 unsigned long addr, unsigned long end)
369 {
370         VMA_ITERATOR(vmi, mm, addr);
371         struct vm_area_struct *vma;
372         unsigned long nr_pages = 0;
373
374         for_each_vma_range(vmi, vma, end) {
375                 unsigned long vm_start = max(addr, vma->vm_start);
376                 unsigned long vm_end = min(end, vma->vm_end);
377
378                 nr_pages += PHYS_PFN(vm_end - vm_start);
379         }
380
381         return nr_pages;
382 }
383
384 static void __vma_link_file(struct vm_area_struct *vma,
385                             struct address_space *mapping)
386 {
387         if (vma_is_shared_maywrite(vma))
388                 mapping_allow_writable(mapping);
389
390         flush_dcache_mmap_lock(mapping);
391         vma_interval_tree_insert(vma, &mapping->i_mmap);
392         flush_dcache_mmap_unlock(mapping);
393 }
394
395 static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
396 {
397         VMA_ITERATOR(vmi, mm, 0);
398         struct address_space *mapping = NULL;
399
400         vma_iter_config(&vmi, vma->vm_start, vma->vm_end);
401         if (vma_iter_prealloc(&vmi, vma))
402                 return -ENOMEM;
403
404         vma_start_write(vma);
405
406         vma_iter_store(&vmi, vma);
407
408         if (vma->vm_file) {
409                 mapping = vma->vm_file->f_mapping;
410                 i_mmap_lock_write(mapping);
411                 __vma_link_file(vma, mapping);
412                 i_mmap_unlock_write(mapping);
413         }
414
415         mm->map_count++;
416         validate_mm(mm);
417         return 0;
418 }
419
420 /*
421  * init_multi_vma_prep() - Initializer for struct vma_prepare
422  * @vp: The vma_prepare struct
423  * @vma: The vma that will be altered once locked
424  * @next: The next vma if it is to be adjusted
425  * @remove: The first vma to be removed
426  * @remove2: The second vma to be removed
427  */
428 static inline void init_multi_vma_prep(struct vma_prepare *vp,
429                 struct vm_area_struct *vma, struct vm_area_struct *next,
430                 struct vm_area_struct *remove, struct vm_area_struct *remove2)
431 {
432         memset(vp, 0, sizeof(struct vma_prepare));
433         vp->vma = vma;
434         vp->anon_vma = vma->anon_vma;
435         vp->remove = remove;
436         vp->remove2 = remove2;
437         vp->adj_next = next;
438         if (!vp->anon_vma && next)
439                 vp->anon_vma = next->anon_vma;
440
441         vp->file = vma->vm_file;
442         if (vp->file)
443                 vp->mapping = vma->vm_file->f_mapping;
444
445 }
446
447 /*
448  * init_vma_prep() - Initializer wrapper for vma_prepare struct
449  * @vp: The vma_prepare struct
450  * @vma: The vma that will be altered once locked
451  */
452 static inline void init_vma_prep(struct vma_prepare *vp,
453                                  struct vm_area_struct *vma)
454 {
455         init_multi_vma_prep(vp, vma, NULL, NULL, NULL);
456 }
457
458
459 /*
460  * vma_prepare() - Helper function for handling locking VMAs prior to altering
461  * @vp: The initialized vma_prepare struct
462  */
463 static inline void vma_prepare(struct vma_prepare *vp)
464 {
465         if (vp->file) {
466                 uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end);
467
468                 if (vp->adj_next)
469                         uprobe_munmap(vp->adj_next, vp->adj_next->vm_start,
470                                       vp->adj_next->vm_end);
471
472                 i_mmap_lock_write(vp->mapping);
473                 if (vp->insert && vp->insert->vm_file) {
474                         /*
475                          * Put into interval tree now, so instantiated pages
476                          * are visible to arm/parisc __flush_dcache_page
477                          * throughout; but we cannot insert into address
478                          * space until vma start or end is updated.
479                          */
480                         __vma_link_file(vp->insert,
481                                         vp->insert->vm_file->f_mapping);
482                 }
483         }
484
485         if (vp->anon_vma) {
486                 anon_vma_lock_write(vp->anon_vma);
487                 anon_vma_interval_tree_pre_update_vma(vp->vma);
488                 if (vp->adj_next)
489                         anon_vma_interval_tree_pre_update_vma(vp->adj_next);
490         }
491
492         if (vp->file) {
493                 flush_dcache_mmap_lock(vp->mapping);
494                 vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap);
495                 if (vp->adj_next)
496                         vma_interval_tree_remove(vp->adj_next,
497                                                  &vp->mapping->i_mmap);
498         }
499
500 }
501
502 /*
503  * vma_complete- Helper function for handling the unlocking after altering VMAs,
504  * or for inserting a VMA.
505  *
506  * @vp: The vma_prepare struct
507  * @vmi: The vma iterator
508  * @mm: The mm_struct
509  */
510 static inline void vma_complete(struct vma_prepare *vp,
511                                 struct vma_iterator *vmi, struct mm_struct *mm)
512 {
513         if (vp->file) {
514                 if (vp->adj_next)
515                         vma_interval_tree_insert(vp->adj_next,
516                                                  &vp->mapping->i_mmap);
517                 vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap);
518                 flush_dcache_mmap_unlock(vp->mapping);
519         }
520
521         if (vp->remove && vp->file) {
522                 __remove_shared_vm_struct(vp->remove, vp->file, vp->mapping);
523                 if (vp->remove2)
524                         __remove_shared_vm_struct(vp->remove2, vp->file,
525                                                   vp->mapping);
526         } else if (vp->insert) {
527                 /*
528                  * split_vma has split insert from vma, and needs
529                  * us to insert it before dropping the locks
530                  * (it may either follow vma or precede it).
531                  */
532                 vma_iter_store(vmi, vp->insert);
533                 mm->map_count++;
534         }
535
536         if (vp->anon_vma) {
537                 anon_vma_interval_tree_post_update_vma(vp->vma);
538                 if (vp->adj_next)
539                         anon_vma_interval_tree_post_update_vma(vp->adj_next);
540                 anon_vma_unlock_write(vp->anon_vma);
541         }
542
543         if (vp->file) {
544                 i_mmap_unlock_write(vp->mapping);
545                 uprobe_mmap(vp->vma);
546
547                 if (vp->adj_next)
548                         uprobe_mmap(vp->adj_next);
549         }
550
551         if (vp->remove) {
552 again:
553                 vma_mark_detached(vp->remove, true);
554                 if (vp->file) {
555                         uprobe_munmap(vp->remove, vp->remove->vm_start,
556                                       vp->remove->vm_end);
557                         fput(vp->file);
558                 }
559                 if (vp->remove->anon_vma)
560                         anon_vma_merge(vp->vma, vp->remove);
561                 mm->map_count--;
562                 mpol_put(vma_policy(vp->remove));
563                 if (!vp->remove2)
564                         WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end);
565                 vm_area_free(vp->remove);
566
567                 /*
568                  * In mprotect's case 6 (see comments on vma_merge),
569                  * we are removing both mid and next vmas
570                  */
571                 if (vp->remove2) {
572                         vp->remove = vp->remove2;
573                         vp->remove2 = NULL;
574                         goto again;
575                 }
576         }
577         if (vp->insert && vp->file)
578                 uprobe_mmap(vp->insert);
579         validate_mm(mm);
580 }
581
582 /*
583  * dup_anon_vma() - Helper function to duplicate anon_vma
584  * @dst: The destination VMA
585  * @src: The source VMA
586  * @dup: Pointer to the destination VMA when successful.
587  *
588  * Returns: 0 on success.
589  */
590 static inline int dup_anon_vma(struct vm_area_struct *dst,
591                 struct vm_area_struct *src, struct vm_area_struct **dup)
592 {
593         /*
594          * Easily overlooked: when mprotect shifts the boundary, make sure the
595          * expanding vma has anon_vma set if the shrinking vma had, to cover any
596          * anon pages imported.
597          */
598         if (src->anon_vma && !dst->anon_vma) {
599                 int ret;
600
601                 vma_assert_write_locked(dst);
602                 dst->anon_vma = src->anon_vma;
603                 ret = anon_vma_clone(dst, src);
604                 if (ret)
605                         return ret;
606
607                 *dup = dst;
608         }
609
610         return 0;
611 }
612
613 /*
614  * vma_expand - Expand an existing VMA
615  *
616  * @vmi: The vma iterator
617  * @vma: The vma to expand
618  * @start: The start of the vma
619  * @end: The exclusive end of the vma
620  * @pgoff: The page offset of vma
621  * @next: The current of next vma.
622  *
623  * Expand @vma to @start and @end.  Can expand off the start and end.  Will
624  * expand over @next if it's different from @vma and @end == @next->vm_end.
625  * Checking if the @vma can expand and merge with @next needs to be handled by
626  * the caller.
627  *
628  * Returns: 0 on success
629  */
630 int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
631                unsigned long start, unsigned long end, pgoff_t pgoff,
632                struct vm_area_struct *next)
633 {
634         struct vm_area_struct *anon_dup = NULL;
635         bool remove_next = false;
636         struct vma_prepare vp;
637
638         vma_start_write(vma);
639         if (next && (vma != next) && (end == next->vm_end)) {
640                 int ret;
641
642                 remove_next = true;
643                 vma_start_write(next);
644                 ret = dup_anon_vma(vma, next, &anon_dup);
645                 if (ret)
646                         return ret;
647         }
648
649         init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
650         /* Not merging but overwriting any part of next is not handled. */
651         VM_WARN_ON(next && !vp.remove &&
652                   next != vma && end > next->vm_start);
653         /* Only handles expanding */
654         VM_WARN_ON(vma->vm_start < start || vma->vm_end > end);
655
656         /* Note: vma iterator must be pointing to 'start' */
657         vma_iter_config(vmi, start, end);
658         if (vma_iter_prealloc(vmi, vma))
659                 goto nomem;
660
661         vma_prepare(&vp);
662         vma_adjust_trans_huge(vma, start, end, 0);
663         vma->vm_start = start;
664         vma->vm_end = end;
665         vma->vm_pgoff = pgoff;
666         vma_iter_store(vmi, vma);
667
668         vma_complete(&vp, vmi, vma->vm_mm);
669         return 0;
670
671 nomem:
672         if (anon_dup)
673                 unlink_anon_vmas(anon_dup);
674         return -ENOMEM;
675 }
676
677 /*
678  * vma_shrink() - Reduce an existing VMAs memory area
679  * @vmi: The vma iterator
680  * @vma: The VMA to modify
681  * @start: The new start
682  * @end: The new end
683  *
684  * Returns: 0 on success, -ENOMEM otherwise
685  */
686 int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
687                unsigned long start, unsigned long end, pgoff_t pgoff)
688 {
689         struct vma_prepare vp;
690
691         WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
692
693         if (vma->vm_start < start)
694                 vma_iter_config(vmi, vma->vm_start, start);
695         else
696                 vma_iter_config(vmi, end, vma->vm_end);
697
698         if (vma_iter_prealloc(vmi, NULL))
699                 return -ENOMEM;
700
701         vma_start_write(vma);
702
703         init_vma_prep(&vp, vma);
704         vma_prepare(&vp);
705         vma_adjust_trans_huge(vma, start, end, 0);
706
707         vma_iter_clear(vmi);
708         vma->vm_start = start;
709         vma->vm_end = end;
710         vma->vm_pgoff = pgoff;
711         vma_complete(&vp, vmi, vma->vm_mm);
712         return 0;
713 }
714
715 /*
716  * If the vma has a ->close operation then the driver probably needs to release
717  * per-vma resources, so we don't attempt to merge those if the caller indicates
718  * the current vma may be removed as part of the merge.
719  */
720 static inline bool is_mergeable_vma(struct vm_area_struct *vma,
721                 struct file *file, unsigned long vm_flags,
722                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
723                 struct anon_vma_name *anon_name, bool may_remove_vma)
724 {
725         /*
726          * VM_SOFTDIRTY should not prevent from VMA merging, if we
727          * match the flags but dirty bit -- the caller should mark
728          * merged VMA as dirty. If dirty bit won't be excluded from
729          * comparison, we increase pressure on the memory system forcing
730          * the kernel to generate new VMAs when old one could be
731          * extended instead.
732          */
733         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
734                 return false;
735         if (vma->vm_file != file)
736                 return false;
737         if (may_remove_vma && vma->vm_ops && vma->vm_ops->close)
738                 return false;
739         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
740                 return false;
741         if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
742                 return false;
743         return true;
744 }
745
746 static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1,
747                  struct anon_vma *anon_vma2, struct vm_area_struct *vma)
748 {
749         /*
750          * The list_is_singular() test is to avoid merging VMA cloned from
751          * parents. This can improve scalability caused by anon_vma lock.
752          */
753         if ((!anon_vma1 || !anon_vma2) && (!vma ||
754                 list_is_singular(&vma->anon_vma_chain)))
755                 return true;
756         return anon_vma1 == anon_vma2;
757 }
758
759 /*
760  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
761  * in front of (at a lower virtual address and file offset than) the vma.
762  *
763  * We cannot merge two vmas if they have differently assigned (non-NULL)
764  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
765  *
766  * We don't check here for the merged mmap wrapping around the end of pagecache
767  * indices (16TB on ia32) because do_mmap() does not permit mmap's which
768  * wrap, nor mmaps which cover the final page at index -1UL.
769  *
770  * We assume the vma may be removed as part of the merge.
771  */
772 static bool
773 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
774                 struct anon_vma *anon_vma, struct file *file,
775                 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
776                 struct anon_vma_name *anon_name)
777 {
778         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, true) &&
779             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
780                 if (vma->vm_pgoff == vm_pgoff)
781                         return true;
782         }
783         return false;
784 }
785
786 /*
787  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
788  * beyond (at a higher virtual address and file offset than) the vma.
789  *
790  * We cannot merge two vmas if they have differently assigned (non-NULL)
791  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
792  *
793  * We assume that vma is not removed as part of the merge.
794  */
795 static bool
796 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
797                 struct anon_vma *anon_vma, struct file *file,
798                 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
799                 struct anon_vma_name *anon_name)
800 {
801         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, false) &&
802             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
803                 pgoff_t vm_pglen;
804                 vm_pglen = vma_pages(vma);
805                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
806                         return true;
807         }
808         return false;
809 }
810
811 /*
812  * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
813  * figure out whether that can be merged with its predecessor or its
814  * successor.  Or both (it neatly fills a hole).
815  *
816  * In most cases - when called for mmap, brk or mremap - [addr,end) is
817  * certain not to be mapped by the time vma_merge is called; but when
818  * called for mprotect, it is certain to be already mapped (either at
819  * an offset within prev, or at the start of next), and the flags of
820  * this area are about to be changed to vm_flags - and the no-change
821  * case has already been eliminated.
822  *
823  * The following mprotect cases have to be considered, where **** is
824  * the area passed down from mprotect_fixup, never extending beyond one
825  * vma, PPPP is the previous vma, CCCC is a concurrent vma that starts
826  * at the same address as **** and is of the same or larger span, and
827  * NNNN the next vma after ****:
828  *
829  *     ****             ****                   ****
830  *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPCCCCCC
831  *    cannot merge    might become       might become
832  *                    PPNNNNNNNNNN       PPPPPPPPPPCC
833  *    mmap, brk or    case 4 below       case 5 below
834  *    mremap move:
835  *                        ****               ****
836  *                    PPPP    NNNN       PPPPCCCCNNNN
837  *                    might become       might become
838  *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
839  *                    PPPPPPPPNNNN 2 or  PPPPPPPPNNNN 7 or
840  *                    PPPPNNNNNNNN 3     PPPPNNNNNNNN 8
841  *
842  * It is important for case 8 that the vma CCCC overlapping the
843  * region **** is never going to extended over NNNN. Instead NNNN must
844  * be extended in region **** and CCCC must be removed. This way in
845  * all cases where vma_merge succeeds, the moment vma_merge drops the
846  * rmap_locks, the properties of the merged vma will be already
847  * correct for the whole merged range. Some of those properties like
848  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
849  * be correct for the whole merged range immediately after the
850  * rmap_locks are released. Otherwise if NNNN would be removed and
851  * CCCC would be extended over the NNNN range, remove_migration_ptes
852  * or other rmap walkers (if working on addresses beyond the "end"
853  * parameter) may establish ptes with the wrong permissions of CCCC
854  * instead of the right permissions of NNNN.
855  *
856  * In the code below:
857  * PPPP is represented by *prev
858  * CCCC is represented by *curr or not represented at all (NULL)
859  * NNNN is represented by *next or not represented at all (NULL)
860  * **** is not represented - it will be merged and the vma containing the
861  *      area is returned, or the function will return NULL
862  */
863 static struct vm_area_struct
864 *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
865            struct vm_area_struct *prev, unsigned long addr, unsigned long end,
866            unsigned long vm_flags, struct anon_vma *anon_vma, struct file *file,
867            pgoff_t pgoff, struct mempolicy *policy,
868            struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
869            struct anon_vma_name *anon_name)
870 {
871         struct vm_area_struct *curr, *next, *res;
872         struct vm_area_struct *vma, *adjust, *remove, *remove2;
873         struct vm_area_struct *anon_dup = NULL;
874         struct vma_prepare vp;
875         pgoff_t vma_pgoff;
876         int err = 0;
877         bool merge_prev = false;
878         bool merge_next = false;
879         bool vma_expanded = false;
880         unsigned long vma_start = addr;
881         unsigned long vma_end = end;
882         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
883         long adj_start = 0;
884
885         /*
886          * We later require that vma->vm_flags == vm_flags,
887          * so this tests vma->vm_flags & VM_SPECIAL, too.
888          */
889         if (vm_flags & VM_SPECIAL)
890                 return NULL;
891
892         /* Does the input range span an existing VMA? (cases 5 - 8) */
893         curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end);
894
895         if (!curr ||                    /* cases 1 - 4 */
896             end == curr->vm_end)        /* cases 6 - 8, adjacent VMA */
897                 next = vma_lookup(mm, end);
898         else
899                 next = NULL;            /* case 5 */
900
901         if (prev) {
902                 vma_start = prev->vm_start;
903                 vma_pgoff = prev->vm_pgoff;
904
905                 /* Can we merge the predecessor? */
906                 if (addr == prev->vm_end && mpol_equal(vma_policy(prev), policy)
907                     && can_vma_merge_after(prev, vm_flags, anon_vma, file,
908                                            pgoff, vm_userfaultfd_ctx, anon_name)) {
909                         merge_prev = true;
910                         vma_prev(vmi);
911                 }
912         }
913
914         /* Can we merge the successor? */
915         if (next && mpol_equal(policy, vma_policy(next)) &&
916             can_vma_merge_before(next, vm_flags, anon_vma, file, pgoff+pglen,
917                                  vm_userfaultfd_ctx, anon_name)) {
918                 merge_next = true;
919         }
920
921         /* Verify some invariant that must be enforced by the caller. */
922         VM_WARN_ON(prev && addr <= prev->vm_start);
923         VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
924         VM_WARN_ON(addr >= end);
925
926         if (!merge_prev && !merge_next)
927                 return NULL; /* Not mergeable. */
928
929         if (merge_prev)
930                 vma_start_write(prev);
931
932         res = vma = prev;
933         remove = remove2 = adjust = NULL;
934
935         /* Can we merge both the predecessor and the successor? */
936         if (merge_prev && merge_next &&
937             is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) {
938                 vma_start_write(next);
939                 remove = next;                          /* case 1 */
940                 vma_end = next->vm_end;
941                 err = dup_anon_vma(prev, next, &anon_dup);
942                 if (curr) {                             /* case 6 */
943                         vma_start_write(curr);
944                         remove = curr;
945                         remove2 = next;
946                         /*
947                          * Note that the dup_anon_vma below cannot overwrite err
948                          * since the first caller would do nothing unless next
949                          * has an anon_vma.
950                          */
951                         if (!next->anon_vma)
952                                 err = dup_anon_vma(prev, curr, &anon_dup);
953                 }
954         } else if (merge_prev) {                        /* case 2 */
955                 if (curr) {
956                         vma_start_write(curr);
957                         err = dup_anon_vma(prev, curr, &anon_dup);
958                         if (end == curr->vm_end) {      /* case 7 */
959                                 remove = curr;
960                         } else {                        /* case 5 */
961                                 adjust = curr;
962                                 adj_start = (end - curr->vm_start);
963                         }
964                 }
965         } else { /* merge_next */
966                 vma_start_write(next);
967                 res = next;
968                 if (prev && addr < prev->vm_end) {      /* case 4 */
969                         vma_start_write(prev);
970                         vma_end = addr;
971                         adjust = next;
972                         adj_start = -(prev->vm_end - addr);
973                         err = dup_anon_vma(next, prev, &anon_dup);
974                 } else {
975                         /*
976                          * Note that cases 3 and 8 are the ONLY ones where prev
977                          * is permitted to be (but is not necessarily) NULL.
978                          */
979                         vma = next;                     /* case 3 */
980                         vma_start = addr;
981                         vma_end = next->vm_end;
982                         vma_pgoff = next->vm_pgoff - pglen;
983                         if (curr) {                     /* case 8 */
984                                 vma_pgoff = curr->vm_pgoff;
985                                 vma_start_write(curr);
986                                 remove = curr;
987                                 err = dup_anon_vma(next, curr, &anon_dup);
988                         }
989                 }
990         }
991
992         /* Error in anon_vma clone. */
993         if (err)
994                 goto anon_vma_fail;
995
996         if (vma_start < vma->vm_start || vma_end > vma->vm_end)
997                 vma_expanded = true;
998
999         if (vma_expanded) {
1000                 vma_iter_config(vmi, vma_start, vma_end);
1001         } else {
1002                 vma_iter_config(vmi, adjust->vm_start + adj_start,
1003                                 adjust->vm_end);
1004         }
1005
1006         if (vma_iter_prealloc(vmi, vma))
1007                 goto prealloc_fail;
1008
1009         init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
1010         VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
1011                    vp.anon_vma != adjust->anon_vma);
1012
1013         vma_prepare(&vp);
1014         vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1015
1016         vma->vm_start = vma_start;
1017         vma->vm_end = vma_end;
1018         vma->vm_pgoff = vma_pgoff;
1019
1020         if (vma_expanded)
1021                 vma_iter_store(vmi, vma);
1022
1023         if (adj_start) {
1024                 adjust->vm_start += adj_start;
1025                 adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1026                 if (adj_start < 0) {
1027                         WARN_ON(vma_expanded);
1028                         vma_iter_store(vmi, next);
1029                 }
1030         }
1031
1032         vma_complete(&vp, vmi, mm);
1033         khugepaged_enter_vma(res, vm_flags);
1034         return res;
1035
1036 prealloc_fail:
1037         if (anon_dup)
1038                 unlink_anon_vmas(anon_dup);
1039
1040 anon_vma_fail:
1041         vma_iter_set(vmi, addr);
1042         vma_iter_load(vmi);
1043         return NULL;
1044 }
1045
1046 /*
1047  * Rough compatibility check to quickly see if it's even worth looking
1048  * at sharing an anon_vma.
1049  *
1050  * They need to have the same vm_file, and the flags can only differ
1051  * in things that mprotect may change.
1052  *
1053  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1054  * we can merge the two vma's. For example, we refuse to merge a vma if
1055  * there is a vm_ops->close() function, because that indicates that the
1056  * driver is doing some kind of reference counting. But that doesn't
1057  * really matter for the anon_vma sharing case.
1058  */
1059 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1060 {
1061         return a->vm_end == b->vm_start &&
1062                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1063                 a->vm_file == b->vm_file &&
1064                 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1065                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1066 }
1067
1068 /*
1069  * Do some basic sanity checking to see if we can re-use the anon_vma
1070  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1071  * the same as 'old', the other will be the new one that is trying
1072  * to share the anon_vma.
1073  *
1074  * NOTE! This runs with mmap_lock held for reading, so it is possible that
1075  * the anon_vma of 'old' is concurrently in the process of being set up
1076  * by another page fault trying to merge _that_. But that's ok: if it
1077  * is being set up, that automatically means that it will be a singleton
1078  * acceptable for merging, so we can do all of this optimistically. But
1079  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1080  *
1081  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1082  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1083  * is to return an anon_vma that is "complex" due to having gone through
1084  * a fork).
1085  *
1086  * We also make sure that the two vma's are compatible (adjacent,
1087  * and with the same memory policies). That's all stable, even with just
1088  * a read lock on the mmap_lock.
1089  */
1090 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1091 {
1092         if (anon_vma_compatible(a, b)) {
1093                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1094
1095                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1096                         return anon_vma;
1097         }
1098         return NULL;
1099 }
1100
1101 /*
1102  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1103  * neighbouring vmas for a suitable anon_vma, before it goes off
1104  * to allocate a new anon_vma.  It checks because a repetitive
1105  * sequence of mprotects and faults may otherwise lead to distinct
1106  * anon_vmas being allocated, preventing vma merge in subsequent
1107  * mprotect.
1108  */
1109 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1110 {
1111         MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1112         struct anon_vma *anon_vma = NULL;
1113         struct vm_area_struct *prev, *next;
1114
1115         /* Try next first. */
1116         next = mas_walk(&mas);
1117         if (next) {
1118                 anon_vma = reusable_anon_vma(next, vma, next);
1119                 if (anon_vma)
1120                         return anon_vma;
1121         }
1122
1123         prev = mas_prev(&mas, 0);
1124         VM_BUG_ON_VMA(prev != vma, vma);
1125         prev = mas_prev(&mas, 0);
1126         /* Try prev next. */
1127         if (prev)
1128                 anon_vma = reusable_anon_vma(prev, prev, vma);
1129
1130         /*
1131          * We might reach here with anon_vma == NULL if we can't find
1132          * any reusable anon_vma.
1133          * There's no absolute need to look only at touching neighbours:
1134          * we could search further afield for "compatible" anon_vmas.
1135          * But it would probably just be a waste of time searching,
1136          * or lead to too many vmas hanging off the same anon_vma.
1137          * We're trying to allow mprotect remerging later on,
1138          * not trying to minimize memory used for anon_vmas.
1139          */
1140         return anon_vma;
1141 }
1142
1143 /*
1144  * If a hint addr is less than mmap_min_addr change hint to be as
1145  * low as possible but still greater than mmap_min_addr
1146  */
1147 static inline unsigned long round_hint_to_min(unsigned long hint)
1148 {
1149         hint &= PAGE_MASK;
1150         if (((void *)hint != NULL) &&
1151             (hint < mmap_min_addr))
1152                 return PAGE_ALIGN(mmap_min_addr);
1153         return hint;
1154 }
1155
1156 bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
1157                         unsigned long bytes)
1158 {
1159         unsigned long locked_pages, limit_pages;
1160
1161         if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK))
1162                 return true;
1163
1164         locked_pages = bytes >> PAGE_SHIFT;
1165         locked_pages += mm->locked_vm;
1166
1167         limit_pages = rlimit(RLIMIT_MEMLOCK);
1168         limit_pages >>= PAGE_SHIFT;
1169
1170         return locked_pages <= limit_pages;
1171 }
1172
1173 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1174 {
1175         if (S_ISREG(inode->i_mode))
1176                 return MAX_LFS_FILESIZE;
1177
1178         if (S_ISBLK(inode->i_mode))
1179                 return MAX_LFS_FILESIZE;
1180
1181         if (S_ISSOCK(inode->i_mode))
1182                 return MAX_LFS_FILESIZE;
1183
1184         /* Special "we do even unsigned file positions" case */
1185         if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1186                 return 0;
1187
1188         /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1189         return ULONG_MAX;
1190 }
1191
1192 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1193                                 unsigned long pgoff, unsigned long len)
1194 {
1195         u64 maxsize = file_mmap_size_max(file, inode);
1196
1197         if (maxsize && len > maxsize)
1198                 return false;
1199         maxsize -= len;
1200         if (pgoff > maxsize >> PAGE_SHIFT)
1201                 return false;
1202         return true;
1203 }
1204
1205 /*
1206  * The caller must write-lock current->mm->mmap_lock.
1207  */
1208 unsigned long do_mmap(struct file *file, unsigned long addr,
1209                         unsigned long len, unsigned long prot,
1210                         unsigned long flags, vm_flags_t vm_flags,
1211                         unsigned long pgoff, unsigned long *populate,
1212                         struct list_head *uf)
1213 {
1214         struct mm_struct *mm = current->mm;
1215         int pkey = 0;
1216
1217         *populate = 0;
1218
1219         if (!len)
1220                 return -EINVAL;
1221
1222         /*
1223          * Does the application expect PROT_READ to imply PROT_EXEC?
1224          *
1225          * (the exception is when the underlying filesystem is noexec
1226          *  mounted, in which case we don't add PROT_EXEC.)
1227          */
1228         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1229                 if (!(file && path_noexec(&file->f_path)))
1230                         prot |= PROT_EXEC;
1231
1232         /* force arch specific MAP_FIXED handling in get_unmapped_area */
1233         if (flags & MAP_FIXED_NOREPLACE)
1234                 flags |= MAP_FIXED;
1235
1236         if (!(flags & MAP_FIXED))
1237                 addr = round_hint_to_min(addr);
1238
1239         /* Careful about overflows.. */
1240         len = PAGE_ALIGN(len);
1241         if (!len)
1242                 return -ENOMEM;
1243
1244         /* offset overflow? */
1245         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1246                 return -EOVERFLOW;
1247
1248         /* Too many mappings? */
1249         if (mm->map_count > sysctl_max_map_count)
1250                 return -ENOMEM;
1251
1252         /* Obtain the address to map to. we verify (or select) it and ensure
1253          * that it represents a valid section of the address space.
1254          */
1255         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1256         if (IS_ERR_VALUE(addr))
1257                 return addr;
1258
1259         if (flags & MAP_FIXED_NOREPLACE) {
1260                 if (find_vma_intersection(mm, addr, addr + len))
1261                         return -EEXIST;
1262         }
1263
1264         if (prot == PROT_EXEC) {
1265                 pkey = execute_only_pkey(mm);
1266                 if (pkey < 0)
1267                         pkey = 0;
1268         }
1269
1270         /* Do simple checking here so the lower-level routines won't have
1271          * to. we assume access permissions have been handled by the open
1272          * of the memory object, so we don't do any here.
1273          */
1274         vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1275                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1276
1277         if (flags & MAP_LOCKED)
1278                 if (!can_do_mlock())
1279                         return -EPERM;
1280
1281         if (!mlock_future_ok(mm, vm_flags, len))
1282                 return -EAGAIN;
1283
1284         if (file) {
1285                 struct inode *inode = file_inode(file);
1286                 unsigned long flags_mask;
1287
1288                 if (!file_mmap_ok(file, inode, pgoff, len))
1289                         return -EOVERFLOW;
1290
1291                 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1292
1293                 switch (flags & MAP_TYPE) {
1294                 case MAP_SHARED:
1295                         /*
1296                          * Force use of MAP_SHARED_VALIDATE with non-legacy
1297                          * flags. E.g. MAP_SYNC is dangerous to use with
1298                          * MAP_SHARED as you don't know which consistency model
1299                          * you will get. We silently ignore unsupported flags
1300                          * with MAP_SHARED to preserve backward compatibility.
1301                          */
1302                         flags &= LEGACY_MAP_MASK;
1303                         fallthrough;
1304                 case MAP_SHARED_VALIDATE:
1305                         if (flags & ~flags_mask)
1306                                 return -EOPNOTSUPP;
1307                         if (prot & PROT_WRITE) {
1308                                 if (!(file->f_mode & FMODE_WRITE))
1309                                         return -EACCES;
1310                                 if (IS_SWAPFILE(file->f_mapping->host))
1311                                         return -ETXTBSY;
1312                         }
1313
1314                         /*
1315                          * Make sure we don't allow writing to an append-only
1316                          * file..
1317                          */
1318                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1319                                 return -EACCES;
1320
1321                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1322                         if (!(file->f_mode & FMODE_WRITE))
1323                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1324                         fallthrough;
1325                 case MAP_PRIVATE:
1326                         if (!(file->f_mode & FMODE_READ))
1327                                 return -EACCES;
1328                         if (path_noexec(&file->f_path)) {
1329                                 if (vm_flags & VM_EXEC)
1330                                         return -EPERM;
1331                                 vm_flags &= ~VM_MAYEXEC;
1332                         }
1333
1334                         if (!file->f_op->mmap)
1335                                 return -ENODEV;
1336                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1337                                 return -EINVAL;
1338                         break;
1339
1340                 default:
1341                         return -EINVAL;
1342                 }
1343         } else {
1344                 switch (flags & MAP_TYPE) {
1345                 case MAP_SHARED:
1346                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1347                                 return -EINVAL;
1348                         /*
1349                          * Ignore pgoff.
1350                          */
1351                         pgoff = 0;
1352                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1353                         break;
1354                 case MAP_PRIVATE:
1355                         /*
1356                          * Set pgoff according to addr for anon_vma.
1357                          */
1358                         pgoff = addr >> PAGE_SHIFT;
1359                         break;
1360                 default:
1361                         return -EINVAL;
1362                 }
1363         }
1364
1365         /*
1366          * Set 'VM_NORESERVE' if we should not account for the
1367          * memory use of this mapping.
1368          */
1369         if (flags & MAP_NORESERVE) {
1370                 /* We honor MAP_NORESERVE if allowed to overcommit */
1371                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1372                         vm_flags |= VM_NORESERVE;
1373
1374                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1375                 if (file && is_file_hugepages(file))
1376                         vm_flags |= VM_NORESERVE;
1377         }
1378
1379         addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1380         if (!IS_ERR_VALUE(addr) &&
1381             ((vm_flags & VM_LOCKED) ||
1382              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1383                 *populate = len;
1384         return addr;
1385 }
1386
1387 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1388                               unsigned long prot, unsigned long flags,
1389                               unsigned long fd, unsigned long pgoff)
1390 {
1391         struct file *file = NULL;
1392         unsigned long retval;
1393
1394         if (!(flags & MAP_ANONYMOUS)) {
1395                 audit_mmap_fd(fd, flags);
1396                 file = fget(fd);
1397                 if (!file)
1398                         return -EBADF;
1399                 if (is_file_hugepages(file)) {
1400                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1401                 } else if (unlikely(flags & MAP_HUGETLB)) {
1402                         retval = -EINVAL;
1403                         goto out_fput;
1404                 }
1405         } else if (flags & MAP_HUGETLB) {
1406                 struct hstate *hs;
1407
1408                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1409                 if (!hs)
1410                         return -EINVAL;
1411
1412                 len = ALIGN(len, huge_page_size(hs));
1413                 /*
1414                  * VM_NORESERVE is used because the reservations will be
1415                  * taken when vm_ops->mmap() is called
1416                  */
1417                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1418                                 VM_NORESERVE,
1419                                 HUGETLB_ANONHUGE_INODE,
1420                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1421                 if (IS_ERR(file))
1422                         return PTR_ERR(file);
1423         }
1424
1425         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1426 out_fput:
1427         if (file)
1428                 fput(file);
1429         return retval;
1430 }
1431
1432 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1433                 unsigned long, prot, unsigned long, flags,
1434                 unsigned long, fd, unsigned long, pgoff)
1435 {
1436         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1437 }
1438
1439 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1440 struct mmap_arg_struct {
1441         unsigned long addr;
1442         unsigned long len;
1443         unsigned long prot;
1444         unsigned long flags;
1445         unsigned long fd;
1446         unsigned long offset;
1447 };
1448
1449 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1450 {
1451         struct mmap_arg_struct a;
1452
1453         if (copy_from_user(&a, arg, sizeof(a)))
1454                 return -EFAULT;
1455         if (offset_in_page(a.offset))
1456                 return -EINVAL;
1457
1458         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1459                                a.offset >> PAGE_SHIFT);
1460 }
1461 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1462
1463 static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops)
1464 {
1465         return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite);
1466 }
1467
1468 static bool vma_is_shared_writable(struct vm_area_struct *vma)
1469 {
1470         return (vma->vm_flags & (VM_WRITE | VM_SHARED)) ==
1471                 (VM_WRITE | VM_SHARED);
1472 }
1473
1474 static bool vma_fs_can_writeback(struct vm_area_struct *vma)
1475 {
1476         /* No managed pages to writeback. */
1477         if (vma->vm_flags & VM_PFNMAP)
1478                 return false;
1479
1480         return vma->vm_file && vma->vm_file->f_mapping &&
1481                 mapping_can_writeback(vma->vm_file->f_mapping);
1482 }
1483
1484 /*
1485  * Does this VMA require the underlying folios to have their dirty state
1486  * tracked?
1487  */
1488 bool vma_needs_dirty_tracking(struct vm_area_struct *vma)
1489 {
1490         /* Only shared, writable VMAs require dirty tracking. */
1491         if (!vma_is_shared_writable(vma))
1492                 return false;
1493
1494         /* Does the filesystem need to be notified? */
1495         if (vm_ops_needs_writenotify(vma->vm_ops))
1496                 return true;
1497
1498         /*
1499          * Even if the filesystem doesn't indicate a need for writenotify, if it
1500          * can writeback, dirty tracking is still required.
1501          */
1502         return vma_fs_can_writeback(vma);
1503 }
1504
1505 /*
1506  * Some shared mappings will want the pages marked read-only
1507  * to track write events. If so, we'll downgrade vm_page_prot
1508  * to the private version (using protection_map[] without the
1509  * VM_SHARED bit).
1510  */
1511 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1512 {
1513         /* If it was private or non-writable, the write bit is already clear */
1514         if (!vma_is_shared_writable(vma))
1515                 return 0;
1516
1517         /* The backer wishes to know when pages are first written to? */
1518         if (vm_ops_needs_writenotify(vma->vm_ops))
1519                 return 1;
1520
1521         /* The open routine did something to the protections that pgprot_modify
1522          * won't preserve? */
1523         if (pgprot_val(vm_page_prot) !=
1524             pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags)))
1525                 return 0;
1526
1527         /*
1528          * Do we need to track softdirty? hugetlb does not support softdirty
1529          * tracking yet.
1530          */
1531         if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1532                 return 1;
1533
1534         /* Do we need write faults for uffd-wp tracking? */
1535         if (userfaultfd_wp(vma))
1536                 return 1;
1537
1538         /* Can the mapping track the dirty pages? */
1539         return vma_fs_can_writeback(vma);
1540 }
1541
1542 /*
1543  * We account for memory if it's a private writeable mapping,
1544  * not hugepages and VM_NORESERVE wasn't set.
1545  */
1546 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1547 {
1548         /*
1549          * hugetlb has its own accounting separate from the core VM
1550          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1551          */
1552         if (file && is_file_hugepages(file))
1553                 return 0;
1554
1555         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1556 }
1557
1558 /**
1559  * unmapped_area() - Find an area between the low_limit and the high_limit with
1560  * the correct alignment and offset, all from @info. Note: current->mm is used
1561  * for the search.
1562  *
1563  * @info: The unmapped area information including the range [low_limit -
1564  * high_limit), the alignment offset and mask.
1565  *
1566  * Return: A memory address or -ENOMEM.
1567  */
1568 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1569 {
1570         unsigned long length, gap;
1571         unsigned long low_limit, high_limit;
1572         struct vm_area_struct *tmp;
1573
1574         MA_STATE(mas, &current->mm->mm_mt, 0, 0);
1575
1576         /* Adjust search length to account for worst case alignment overhead */
1577         length = info->length + info->align_mask;
1578         if (length < info->length)
1579                 return -ENOMEM;
1580
1581         low_limit = info->low_limit;
1582         if (low_limit < mmap_min_addr)
1583                 low_limit = mmap_min_addr;
1584         high_limit = info->high_limit;
1585 retry:
1586         if (mas_empty_area(&mas, low_limit, high_limit - 1, length))
1587                 return -ENOMEM;
1588
1589         gap = mas.index;
1590         gap += (info->align_offset - gap) & info->align_mask;
1591         tmp = mas_next(&mas, ULONG_MAX);
1592         if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1593                 if (vm_start_gap(tmp) < gap + length - 1) {
1594                         low_limit = tmp->vm_end;
1595                         mas_reset(&mas);
1596                         goto retry;
1597                 }
1598         } else {
1599                 tmp = mas_prev(&mas, 0);
1600                 if (tmp && vm_end_gap(tmp) > gap) {
1601                         low_limit = vm_end_gap(tmp);
1602                         mas_reset(&mas);
1603                         goto retry;
1604                 }
1605         }
1606
1607         return gap;
1608 }
1609
1610 /**
1611  * unmapped_area_topdown() - Find an area between the low_limit and the
1612  * high_limit with the correct alignment and offset at the highest available
1613  * address, all from @info. Note: current->mm is used for the search.
1614  *
1615  * @info: The unmapped area information including the range [low_limit -
1616  * high_limit), the alignment offset and mask.
1617  *
1618  * Return: A memory address or -ENOMEM.
1619  */
1620 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1621 {
1622         unsigned long length, gap, gap_end;
1623         unsigned long low_limit, high_limit;
1624         struct vm_area_struct *tmp;
1625
1626         MA_STATE(mas, &current->mm->mm_mt, 0, 0);
1627         /* Adjust search length to account for worst case alignment overhead */
1628         length = info->length + info->align_mask;
1629         if (length < info->length)
1630                 return -ENOMEM;
1631
1632         low_limit = info->low_limit;
1633         if (low_limit < mmap_min_addr)
1634                 low_limit = mmap_min_addr;
1635         high_limit = info->high_limit;
1636 retry:
1637         if (mas_empty_area_rev(&mas, low_limit, high_limit - 1, length))
1638                 return -ENOMEM;
1639
1640         gap = mas.last + 1 - info->length;
1641         gap -= (gap - info->align_offset) & info->align_mask;
1642         gap_end = mas.last;
1643         tmp = mas_next(&mas, ULONG_MAX);
1644         if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1645                 if (vm_start_gap(tmp) <= gap_end) {
1646                         high_limit = vm_start_gap(tmp);
1647                         mas_reset(&mas);
1648                         goto retry;
1649                 }
1650         } else {
1651                 tmp = mas_prev(&mas, 0);
1652                 if (tmp && vm_end_gap(tmp) > gap) {
1653                         high_limit = tmp->vm_start;
1654                         mas_reset(&mas);
1655                         goto retry;
1656                 }
1657         }
1658
1659         return gap;
1660 }
1661
1662 /*
1663  * Search for an unmapped address range.
1664  *
1665  * We are looking for a range that:
1666  * - does not intersect with any VMA;
1667  * - is contained within the [low_limit, high_limit) interval;
1668  * - is at least the desired size.
1669  * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1670  */
1671 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1672 {
1673         unsigned long addr;
1674
1675         if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1676                 addr = unmapped_area_topdown(info);
1677         else
1678                 addr = unmapped_area(info);
1679
1680         trace_vm_unmapped_area(addr, info);
1681         return addr;
1682 }
1683
1684 /* Get an address range which is currently unmapped.
1685  * For shmat() with addr=0.
1686  *
1687  * Ugly calling convention alert:
1688  * Return value with the low bits set means error value,
1689  * ie
1690  *      if (ret & ~PAGE_MASK)
1691  *              error = ret;
1692  *
1693  * This function "knows" that -ENOMEM has the bits set.
1694  */
1695 unsigned long
1696 generic_get_unmapped_area(struct file *filp, unsigned long addr,
1697                           unsigned long len, unsigned long pgoff,
1698                           unsigned long flags)
1699 {
1700         struct mm_struct *mm = current->mm;
1701         struct vm_area_struct *vma, *prev;
1702         struct vm_unmapped_area_info info;
1703         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1704
1705         if (len > mmap_end - mmap_min_addr)
1706                 return -ENOMEM;
1707
1708         if (flags & MAP_FIXED)
1709                 return addr;
1710
1711         if (addr) {
1712                 addr = PAGE_ALIGN(addr);
1713                 vma = find_vma_prev(mm, addr, &prev);
1714                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1715                     (!vma || addr + len <= vm_start_gap(vma)) &&
1716                     (!prev || addr >= vm_end_gap(prev)))
1717                         return addr;
1718         }
1719
1720         info.flags = 0;
1721         info.length = len;
1722         info.low_limit = mm->mmap_base;
1723         info.high_limit = mmap_end;
1724         info.align_mask = 0;
1725         info.align_offset = 0;
1726         return vm_unmapped_area(&info);
1727 }
1728
1729 #ifndef HAVE_ARCH_UNMAPPED_AREA
1730 unsigned long
1731 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1732                        unsigned long len, unsigned long pgoff,
1733                        unsigned long flags)
1734 {
1735         return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1736 }
1737 #endif
1738
1739 /*
1740  * This mmap-allocator allocates new areas top-down from below the
1741  * stack's low limit (the base):
1742  */
1743 unsigned long
1744 generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1745                                   unsigned long len, unsigned long pgoff,
1746                                   unsigned long flags)
1747 {
1748         struct vm_area_struct *vma, *prev;
1749         struct mm_struct *mm = current->mm;
1750         struct vm_unmapped_area_info info;
1751         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1752
1753         /* requested length too big for entire address space */
1754         if (len > mmap_end - mmap_min_addr)
1755                 return -ENOMEM;
1756
1757         if (flags & MAP_FIXED)
1758                 return addr;
1759
1760         /* requesting a specific address */
1761         if (addr) {
1762                 addr = PAGE_ALIGN(addr);
1763                 vma = find_vma_prev(mm, addr, &prev);
1764                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1765                                 (!vma || addr + len <= vm_start_gap(vma)) &&
1766                                 (!prev || addr >= vm_end_gap(prev)))
1767                         return addr;
1768         }
1769
1770         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1771         info.length = len;
1772         info.low_limit = PAGE_SIZE;
1773         info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1774         info.align_mask = 0;
1775         info.align_offset = 0;
1776         addr = vm_unmapped_area(&info);
1777
1778         /*
1779          * A failed mmap() very likely causes application failure,
1780          * so fall back to the bottom-up function here. This scenario
1781          * can happen with large stack limits and large mmap()
1782          * allocations.
1783          */
1784         if (offset_in_page(addr)) {
1785                 VM_BUG_ON(addr != -ENOMEM);
1786                 info.flags = 0;
1787                 info.low_limit = TASK_UNMAPPED_BASE;
1788                 info.high_limit = mmap_end;
1789                 addr = vm_unmapped_area(&info);
1790         }
1791
1792         return addr;
1793 }
1794
1795 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1796 unsigned long
1797 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1798                                unsigned long len, unsigned long pgoff,
1799                                unsigned long flags)
1800 {
1801         return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1802 }
1803 #endif
1804
1805 unsigned long
1806 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1807                 unsigned long pgoff, unsigned long flags)
1808 {
1809         unsigned long (*get_area)(struct file *, unsigned long,
1810                                   unsigned long, unsigned long, unsigned long);
1811
1812         unsigned long error = arch_mmap_check(addr, len, flags);
1813         if (error)
1814                 return error;
1815
1816         /* Careful about overflows.. */
1817         if (len > TASK_SIZE)
1818                 return -ENOMEM;
1819
1820         get_area = current->mm->get_unmapped_area;
1821         if (file) {
1822                 if (file->f_op->get_unmapped_area)
1823                         get_area = file->f_op->get_unmapped_area;
1824         } else if (flags & MAP_SHARED) {
1825                 /*
1826                  * mmap_region() will call shmem_zero_setup() to create a file,
1827                  * so use shmem's get_unmapped_area in case it can be huge.
1828                  * do_mmap() will clear pgoff, so match alignment.
1829                  */
1830                 pgoff = 0;
1831                 get_area = shmem_get_unmapped_area;
1832         }
1833
1834         addr = get_area(file, addr, len, pgoff, flags);
1835         if (IS_ERR_VALUE(addr))
1836                 return addr;
1837
1838         if (addr > TASK_SIZE - len)
1839                 return -ENOMEM;
1840         if (offset_in_page(addr))
1841                 return -EINVAL;
1842
1843         error = security_mmap_addr(addr);
1844         return error ? error : addr;
1845 }
1846
1847 EXPORT_SYMBOL(get_unmapped_area);
1848
1849 /**
1850  * find_vma_intersection() - Look up the first VMA which intersects the interval
1851  * @mm: The process address space.
1852  * @start_addr: The inclusive start user address.
1853  * @end_addr: The exclusive end user address.
1854  *
1855  * Returns: The first VMA within the provided range, %NULL otherwise.  Assumes
1856  * start_addr < end_addr.
1857  */
1858 struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1859                                              unsigned long start_addr,
1860                                              unsigned long end_addr)
1861 {
1862         unsigned long index = start_addr;
1863
1864         mmap_assert_locked(mm);
1865         return mt_find(&mm->mm_mt, &index, end_addr - 1);
1866 }
1867 EXPORT_SYMBOL(find_vma_intersection);
1868
1869 /**
1870  * find_vma() - Find the VMA for a given address, or the next VMA.
1871  * @mm: The mm_struct to check
1872  * @addr: The address
1873  *
1874  * Returns: The VMA associated with addr, or the next VMA.
1875  * May return %NULL in the case of no VMA at addr or above.
1876  */
1877 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1878 {
1879         unsigned long index = addr;
1880
1881         mmap_assert_locked(mm);
1882         return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1883 }
1884 EXPORT_SYMBOL(find_vma);
1885
1886 /**
1887  * find_vma_prev() - Find the VMA for a given address, or the next vma and
1888  * set %pprev to the previous VMA, if any.
1889  * @mm: The mm_struct to check
1890  * @addr: The address
1891  * @pprev: The pointer to set to the previous VMA
1892  *
1893  * Note that RCU lock is missing here since the external mmap_lock() is used
1894  * instead.
1895  *
1896  * Returns: The VMA associated with @addr, or the next vma.
1897  * May return %NULL in the case of no vma at addr or above.
1898  */
1899 struct vm_area_struct *
1900 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1901                         struct vm_area_struct **pprev)
1902 {
1903         struct vm_area_struct *vma;
1904         MA_STATE(mas, &mm->mm_mt, addr, addr);
1905
1906         vma = mas_walk(&mas);
1907         *pprev = mas_prev(&mas, 0);
1908         if (!vma)
1909                 vma = mas_next(&mas, ULONG_MAX);
1910         return vma;
1911 }
1912
1913 /*
1914  * Verify that the stack growth is acceptable and
1915  * update accounting. This is shared with both the
1916  * grow-up and grow-down cases.
1917  */
1918 static int acct_stack_growth(struct vm_area_struct *vma,
1919                              unsigned long size, unsigned long grow)
1920 {
1921         struct mm_struct *mm = vma->vm_mm;
1922         unsigned long new_start;
1923
1924         /* address space limit tests */
1925         if (!may_expand_vm(mm, vma->vm_flags, grow))
1926                 return -ENOMEM;
1927
1928         /* Stack limit test */
1929         if (size > rlimit(RLIMIT_STACK))
1930                 return -ENOMEM;
1931
1932         /* mlock limit tests */
1933         if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT))
1934                 return -ENOMEM;
1935
1936         /* Check to ensure the stack will not grow into a hugetlb-only region */
1937         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1938                         vma->vm_end - size;
1939         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1940                 return -EFAULT;
1941
1942         /*
1943          * Overcommit..  This must be the final test, as it will
1944          * update security statistics.
1945          */
1946         if (security_vm_enough_memory_mm(mm, grow))
1947                 return -ENOMEM;
1948
1949         return 0;
1950 }
1951
1952 #if defined(CONFIG_STACK_GROWSUP)
1953 /*
1954  * PA-RISC uses this for its stack.
1955  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1956  */
1957 static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1958 {
1959         struct mm_struct *mm = vma->vm_mm;
1960         struct vm_area_struct *next;
1961         unsigned long gap_addr;
1962         int error = 0;
1963         MA_STATE(mas, &mm->mm_mt, vma->vm_start, address);
1964
1965         if (!(vma->vm_flags & VM_GROWSUP))
1966                 return -EFAULT;
1967
1968         /* Guard against exceeding limits of the address space. */
1969         address &= PAGE_MASK;
1970         if (address >= (TASK_SIZE & PAGE_MASK))
1971                 return -ENOMEM;
1972         address += PAGE_SIZE;
1973
1974         /* Enforce stack_guard_gap */
1975         gap_addr = address + stack_guard_gap;
1976
1977         /* Guard against overflow */
1978         if (gap_addr < address || gap_addr > TASK_SIZE)
1979                 gap_addr = TASK_SIZE;
1980
1981         next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1982         if (next && vma_is_accessible(next)) {
1983                 if (!(next->vm_flags & VM_GROWSUP))
1984                         return -ENOMEM;
1985                 /* Check that both stack segments have the same anon_vma? */
1986         }
1987
1988         if (next)
1989                 mas_prev_range(&mas, address);
1990
1991         __mas_set_range(&mas, vma->vm_start, address - 1);
1992         if (mas_preallocate(&mas, vma, GFP_KERNEL))
1993                 return -ENOMEM;
1994
1995         /* We must make sure the anon_vma is allocated. */
1996         if (unlikely(anon_vma_prepare(vma))) {
1997                 mas_destroy(&mas);
1998                 return -ENOMEM;
1999         }
2000
2001         /* Lock the VMA before expanding to prevent concurrent page faults */
2002         vma_start_write(vma);
2003         /*
2004          * vma->vm_start/vm_end cannot change under us because the caller
2005          * is required to hold the mmap_lock in read mode.  We need the
2006          * anon_vma lock to serialize against concurrent expand_stacks.
2007          */
2008         anon_vma_lock_write(vma->anon_vma);
2009
2010         /* Somebody else might have raced and expanded it already */
2011         if (address > vma->vm_end) {
2012                 unsigned long size, grow;
2013
2014                 size = address - vma->vm_start;
2015                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2016
2017                 error = -ENOMEM;
2018                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2019                         error = acct_stack_growth(vma, size, grow);
2020                         if (!error) {
2021                                 /*
2022                                  * We only hold a shared mmap_lock lock here, so
2023                                  * we need to protect against concurrent vma
2024                                  * expansions.  anon_vma_lock_write() doesn't
2025                                  * help here, as we don't guarantee that all
2026                                  * growable vmas in a mm share the same root
2027                                  * anon vma.  So, we reuse mm->page_table_lock
2028                                  * to guard against concurrent vma expansions.
2029                                  */
2030                                 spin_lock(&mm->page_table_lock);
2031                                 if (vma->vm_flags & VM_LOCKED)
2032                                         mm->locked_vm += grow;
2033                                 vm_stat_account(mm, vma->vm_flags, grow);
2034                                 anon_vma_interval_tree_pre_update_vma(vma);
2035                                 vma->vm_end = address;
2036                                 /* Overwrite old entry in mtree. */
2037                                 mas_store_prealloc(&mas, vma);
2038                                 anon_vma_interval_tree_post_update_vma(vma);
2039                                 spin_unlock(&mm->page_table_lock);
2040
2041                                 perf_event_mmap(vma);
2042                         }
2043                 }
2044         }
2045         anon_vma_unlock_write(vma->anon_vma);
2046         khugepaged_enter_vma(vma, vma->vm_flags);
2047         mas_destroy(&mas);
2048         validate_mm(mm);
2049         return error;
2050 }
2051 #endif /* CONFIG_STACK_GROWSUP */
2052
2053 /*
2054  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2055  * mmap_lock held for writing.
2056  */
2057 int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2058 {
2059         struct mm_struct *mm = vma->vm_mm;
2060         MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2061         struct vm_area_struct *prev;
2062         int error = 0;
2063
2064         if (!(vma->vm_flags & VM_GROWSDOWN))
2065                 return -EFAULT;
2066
2067         address &= PAGE_MASK;
2068         if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
2069                 return -EPERM;
2070
2071         /* Enforce stack_guard_gap */
2072         prev = mas_prev(&mas, 0);
2073         /* Check that both stack segments have the same anon_vma? */
2074         if (prev) {
2075                 if (!(prev->vm_flags & VM_GROWSDOWN) &&
2076                     vma_is_accessible(prev) &&
2077                     (address - prev->vm_end < stack_guard_gap))
2078                         return -ENOMEM;
2079         }
2080
2081         if (prev)
2082                 mas_next_range(&mas, vma->vm_start);
2083
2084         __mas_set_range(&mas, address, vma->vm_end - 1);
2085         if (mas_preallocate(&mas, vma, GFP_KERNEL))
2086                 return -ENOMEM;
2087
2088         /* We must make sure the anon_vma is allocated. */
2089         if (unlikely(anon_vma_prepare(vma))) {
2090                 mas_destroy(&mas);
2091                 return -ENOMEM;
2092         }
2093
2094         /* Lock the VMA before expanding to prevent concurrent page faults */
2095         vma_start_write(vma);
2096         /*
2097          * vma->vm_start/vm_end cannot change under us because the caller
2098          * is required to hold the mmap_lock in read mode.  We need the
2099          * anon_vma lock to serialize against concurrent expand_stacks.
2100          */
2101         anon_vma_lock_write(vma->anon_vma);
2102
2103         /* Somebody else might have raced and expanded it already */
2104         if (address < vma->vm_start) {
2105                 unsigned long size, grow;
2106
2107                 size = vma->vm_end - address;
2108                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2109
2110                 error = -ENOMEM;
2111                 if (grow <= vma->vm_pgoff) {
2112                         error = acct_stack_growth(vma, size, grow);
2113                         if (!error) {
2114                                 /*
2115                                  * We only hold a shared mmap_lock lock here, so
2116                                  * we need to protect against concurrent vma
2117                                  * expansions.  anon_vma_lock_write() doesn't
2118                                  * help here, as we don't guarantee that all
2119                                  * growable vmas in a mm share the same root
2120                                  * anon vma.  So, we reuse mm->page_table_lock
2121                                  * to guard against concurrent vma expansions.
2122                                  */
2123                                 spin_lock(&mm->page_table_lock);
2124                                 if (vma->vm_flags & VM_LOCKED)
2125                                         mm->locked_vm += grow;
2126                                 vm_stat_account(mm, vma->vm_flags, grow);
2127                                 anon_vma_interval_tree_pre_update_vma(vma);
2128                                 vma->vm_start = address;
2129                                 vma->vm_pgoff -= grow;
2130                                 /* Overwrite old entry in mtree. */
2131                                 mas_store_prealloc(&mas, vma);
2132                                 anon_vma_interval_tree_post_update_vma(vma);
2133                                 spin_unlock(&mm->page_table_lock);
2134
2135                                 perf_event_mmap(vma);
2136                         }
2137                 }
2138         }
2139         anon_vma_unlock_write(vma->anon_vma);
2140         khugepaged_enter_vma(vma, vma->vm_flags);
2141         mas_destroy(&mas);
2142         validate_mm(mm);
2143         return error;
2144 }
2145
2146 /* enforced gap between the expanding stack and other mappings. */
2147 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2148
2149 static int __init cmdline_parse_stack_guard_gap(char *p)
2150 {
2151         unsigned long val;
2152         char *endptr;
2153
2154         val = simple_strtoul(p, &endptr, 10);
2155         if (!*endptr)
2156                 stack_guard_gap = val << PAGE_SHIFT;
2157
2158         return 1;
2159 }
2160 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2161
2162 #ifdef CONFIG_STACK_GROWSUP
2163 int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2164 {
2165         return expand_upwards(vma, address);
2166 }
2167
2168 struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2169 {
2170         struct vm_area_struct *vma, *prev;
2171
2172         addr &= PAGE_MASK;
2173         vma = find_vma_prev(mm, addr, &prev);
2174         if (vma && (vma->vm_start <= addr))
2175                 return vma;
2176         if (!prev)
2177                 return NULL;
2178         if (expand_stack_locked(prev, addr))
2179                 return NULL;
2180         if (prev->vm_flags & VM_LOCKED)
2181                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2182         return prev;
2183 }
2184 #else
2185 int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2186 {
2187         return expand_downwards(vma, address);
2188 }
2189
2190 struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2191 {
2192         struct vm_area_struct *vma;
2193         unsigned long start;
2194
2195         addr &= PAGE_MASK;
2196         vma = find_vma(mm, addr);
2197         if (!vma)
2198                 return NULL;
2199         if (vma->vm_start <= addr)
2200                 return vma;
2201         start = vma->vm_start;
2202         if (expand_stack_locked(vma, addr))
2203                 return NULL;
2204         if (vma->vm_flags & VM_LOCKED)
2205                 populate_vma_page_range(vma, addr, start, NULL);
2206         return vma;
2207 }
2208 #endif
2209
2210 #if defined(CONFIG_STACK_GROWSUP)
2211
2212 #define vma_expand_up(vma,addr) expand_upwards(vma, addr)
2213 #define vma_expand_down(vma, addr) (-EFAULT)
2214
2215 #else
2216
2217 #define vma_expand_up(vma,addr) (-EFAULT)
2218 #define vma_expand_down(vma, addr) expand_downwards(vma, addr)
2219
2220 #endif
2221
2222 /*
2223  * expand_stack(): legacy interface for page faulting. Don't use unless
2224  * you have to.
2225  *
2226  * This is called with the mm locked for reading, drops the lock, takes
2227  * the lock for writing, tries to look up a vma again, expands it if
2228  * necessary, and downgrades the lock to reading again.
2229  *
2230  * If no vma is found or it can't be expanded, it returns NULL and has
2231  * dropped the lock.
2232  */
2233 struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
2234 {
2235         struct vm_area_struct *vma, *prev;
2236
2237         mmap_read_unlock(mm);
2238         if (mmap_write_lock_killable(mm))
2239                 return NULL;
2240
2241         vma = find_vma_prev(mm, addr, &prev);
2242         if (vma && vma->vm_start <= addr)
2243                 goto success;
2244
2245         if (prev && !vma_expand_up(prev, addr)) {
2246                 vma = prev;
2247                 goto success;
2248         }
2249
2250         if (vma && !vma_expand_down(vma, addr))
2251                 goto success;
2252
2253         mmap_write_unlock(mm);
2254         return NULL;
2255
2256 success:
2257         mmap_write_downgrade(mm);
2258         return vma;
2259 }
2260
2261 /*
2262  * Ok - we have the memory areas we should free on a maple tree so release them,
2263  * and do the vma updates.
2264  *
2265  * Called with the mm semaphore held.
2266  */
2267 static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2268 {
2269         unsigned long nr_accounted = 0;
2270         struct vm_area_struct *vma;
2271
2272         /* Update high watermark before we lower total_vm */
2273         update_hiwater_vm(mm);
2274         mas_for_each(mas, vma, ULONG_MAX) {
2275                 long nrpages = vma_pages(vma);
2276
2277                 if (vma->vm_flags & VM_ACCOUNT)
2278                         nr_accounted += nrpages;
2279                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2280                 remove_vma(vma, false);
2281         }
2282         vm_unacct_memory(nr_accounted);
2283 }
2284
2285 /*
2286  * Get rid of page table information in the indicated region.
2287  *
2288  * Called with the mm semaphore held.
2289  */
2290 static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
2291                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2292                 struct vm_area_struct *next, unsigned long start,
2293                 unsigned long end, unsigned long tree_end, bool mm_wr_locked)
2294 {
2295         struct mmu_gather tlb;
2296         unsigned long mt_start = mas->index;
2297
2298         lru_add_drain();
2299         tlb_gather_mmu(&tlb, mm);
2300         update_hiwater_rss(mm);
2301         unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
2302         mas_set(mas, mt_start);
2303         free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2304                                  next ? next->vm_start : USER_PGTABLES_CEILING,
2305                                  mm_wr_locked);
2306         tlb_finish_mmu(&tlb);
2307 }
2308
2309 /*
2310  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2311  * has already been checked or doesn't make sense to fail.
2312  * VMA Iterator will point to the end VMA.
2313  */
2314 static int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2315                        unsigned long addr, int new_below)
2316 {
2317         struct vma_prepare vp;
2318         struct vm_area_struct *new;
2319         int err;
2320
2321         WARN_ON(vma->vm_start >= addr);
2322         WARN_ON(vma->vm_end <= addr);
2323
2324         if (vma->vm_ops && vma->vm_ops->may_split) {
2325                 err = vma->vm_ops->may_split(vma, addr);
2326                 if (err)
2327                         return err;
2328         }
2329
2330         new = vm_area_dup(vma);
2331         if (!new)
2332                 return -ENOMEM;
2333
2334         if (new_below) {
2335                 new->vm_end = addr;
2336         } else {
2337                 new->vm_start = addr;
2338                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2339         }
2340
2341         err = -ENOMEM;
2342         vma_iter_config(vmi, new->vm_start, new->vm_end);
2343         if (vma_iter_prealloc(vmi, new))
2344                 goto out_free_vma;
2345
2346         err = vma_dup_policy(vma, new);
2347         if (err)
2348                 goto out_free_vmi;
2349
2350         err = anon_vma_clone(new, vma);
2351         if (err)
2352                 goto out_free_mpol;
2353
2354         if (new->vm_file)
2355                 get_file(new->vm_file);
2356
2357         if (new->vm_ops && new->vm_ops->open)
2358                 new->vm_ops->open(new);
2359
2360         vma_start_write(vma);
2361         vma_start_write(new);
2362
2363         init_vma_prep(&vp, vma);
2364         vp.insert = new;
2365         vma_prepare(&vp);
2366         vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
2367
2368         if (new_below) {
2369                 vma->vm_start = addr;
2370                 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
2371         } else {
2372                 vma->vm_end = addr;
2373         }
2374
2375         /* vma_complete stores the new vma */
2376         vma_complete(&vp, vmi, vma->vm_mm);
2377
2378         /* Success. */
2379         if (new_below)
2380                 vma_next(vmi);
2381         return 0;
2382
2383 out_free_mpol:
2384         mpol_put(vma_policy(new));
2385 out_free_vmi:
2386         vma_iter_free(vmi);
2387 out_free_vma:
2388         vm_area_free(new);
2389         return err;
2390 }
2391
2392 /*
2393  * Split a vma into two pieces at address 'addr', a new vma is allocated
2394  * either for the first part or the tail.
2395  */
2396 static int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2397                      unsigned long addr, int new_below)
2398 {
2399         if (vma->vm_mm->map_count >= sysctl_max_map_count)
2400                 return -ENOMEM;
2401
2402         return __split_vma(vmi, vma, addr, new_below);
2403 }
2404
2405 /*
2406  * We are about to modify one or multiple of a VMA's flags, policy, userfaultfd
2407  * context and anonymous VMA name within the range [start, end).
2408  *
2409  * As a result, we might be able to merge the newly modified VMA range with an
2410  * adjacent VMA with identical properties.
2411  *
2412  * If no merge is possible and the range does not span the entirety of the VMA,
2413  * we then need to split the VMA to accommodate the change.
2414  *
2415  * The function returns either the merged VMA, the original VMA if a split was
2416  * required instead, or an error if the split failed.
2417  */
2418 struct vm_area_struct *vma_modify(struct vma_iterator *vmi,
2419                                   struct vm_area_struct *prev,
2420                                   struct vm_area_struct *vma,
2421                                   unsigned long start, unsigned long end,
2422                                   unsigned long vm_flags,
2423                                   struct mempolicy *policy,
2424                                   struct vm_userfaultfd_ctx uffd_ctx,
2425                                   struct anon_vma_name *anon_name)
2426 {
2427         pgoff_t pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
2428         struct vm_area_struct *merged;
2429
2430         merged = vma_merge(vmi, vma->vm_mm, prev, start, end, vm_flags,
2431                            vma->anon_vma, vma->vm_file, pgoff, policy,
2432                            uffd_ctx, anon_name);
2433         if (merged)
2434                 return merged;
2435
2436         if (vma->vm_start < start) {
2437                 int err = split_vma(vmi, vma, start, 1);
2438
2439                 if (err)
2440                         return ERR_PTR(err);
2441         }
2442
2443         if (vma->vm_end > end) {
2444                 int err = split_vma(vmi, vma, end, 0);
2445
2446                 if (err)
2447                         return ERR_PTR(err);
2448         }
2449
2450         return vma;
2451 }
2452
2453 /*
2454  * Attempt to merge a newly mapped VMA with those adjacent to it. The caller
2455  * must ensure that [start, end) does not overlap any existing VMA.
2456  */
2457 static struct vm_area_struct
2458 *vma_merge_new_vma(struct vma_iterator *vmi, struct vm_area_struct *prev,
2459                    struct vm_area_struct *vma, unsigned long start,
2460                    unsigned long end, pgoff_t pgoff)
2461 {
2462         return vma_merge(vmi, vma->vm_mm, prev, start, end, vma->vm_flags,
2463                          vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
2464                          vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2465 }
2466
2467 /*
2468  * Expand vma by delta bytes, potentially merging with an immediately adjacent
2469  * VMA with identical properties.
2470  */
2471 struct vm_area_struct *vma_merge_extend(struct vma_iterator *vmi,
2472                                         struct vm_area_struct *vma,
2473                                         unsigned long delta)
2474 {
2475         pgoff_t pgoff = vma->vm_pgoff + vma_pages(vma);
2476
2477         /* vma is specified as prev, so case 1 or 2 will apply. */
2478         return vma_merge(vmi, vma->vm_mm, vma, vma->vm_end, vma->vm_end + delta,
2479                          vma->vm_flags, vma->anon_vma, vma->vm_file, pgoff,
2480                          vma_policy(vma), vma->vm_userfaultfd_ctx,
2481                          anon_vma_name(vma));
2482 }
2483
2484 /*
2485  * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
2486  * @vmi: The vma iterator
2487  * @vma: The starting vm_area_struct
2488  * @mm: The mm_struct
2489  * @start: The aligned start address to munmap.
2490  * @end: The aligned end address to munmap.
2491  * @uf: The userfaultfd list_head
2492  * @unlock: Set to true to drop the mmap_lock.  unlocking only happens on
2493  * success.
2494  *
2495  * Return: 0 on success and drops the lock if so directed, error and leaves the
2496  * lock held otherwise.
2497  */
2498 static int
2499 do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2500                     struct mm_struct *mm, unsigned long start,
2501                     unsigned long end, struct list_head *uf, bool unlock)
2502 {
2503         struct vm_area_struct *prev, *next = NULL;
2504         struct maple_tree mt_detach;
2505         int count = 0;
2506         int error = -ENOMEM;
2507         unsigned long locked_vm = 0;
2508         MA_STATE(mas_detach, &mt_detach, 0, 0);
2509         mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
2510         mt_on_stack(mt_detach);
2511
2512         /*
2513          * If we need to split any vma, do it now to save pain later.
2514          *
2515          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2516          * unmapped vm_area_struct will remain in use: so lower split_vma
2517          * places tmp vma above, and higher split_vma places tmp vma below.
2518          */
2519
2520         /* Does it split the first one? */
2521         if (start > vma->vm_start) {
2522
2523                 /*
2524                  * Make sure that map_count on return from munmap() will
2525                  * not exceed its limit; but let map_count go just above
2526                  * its limit temporarily, to help free resources as expected.
2527                  */
2528                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2529                         goto map_count_exceeded;
2530
2531                 error = __split_vma(vmi, vma, start, 1);
2532                 if (error)
2533                         goto start_split_failed;
2534         }
2535
2536         /*
2537          * Detach a range of VMAs from the mm. Using next as a temp variable as
2538          * it is always overwritten.
2539          */
2540         next = vma;
2541         do {
2542                 /* Does it split the end? */
2543                 if (next->vm_end > end) {
2544                         error = __split_vma(vmi, next, end, 0);
2545                         if (error)
2546                                 goto end_split_failed;
2547                 }
2548                 vma_start_write(next);
2549                 mas_set(&mas_detach, count);
2550                 error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
2551                 if (error)
2552                         goto munmap_gather_failed;
2553                 vma_mark_detached(next, true);
2554                 if (next->vm_flags & VM_LOCKED)
2555                         locked_vm += vma_pages(next);
2556
2557                 count++;
2558                 if (unlikely(uf)) {
2559                         /*
2560                          * If userfaultfd_unmap_prep returns an error the vmas
2561                          * will remain split, but userland will get a
2562                          * highly unexpected error anyway. This is no
2563                          * different than the case where the first of the two
2564                          * __split_vma fails, but we don't undo the first
2565                          * split, despite we could. This is unlikely enough
2566                          * failure that it's not worth optimizing it for.
2567                          */
2568                         error = userfaultfd_unmap_prep(next, start, end, uf);
2569
2570                         if (error)
2571                                 goto userfaultfd_error;
2572                 }
2573 #ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2574                 BUG_ON(next->vm_start < start);
2575                 BUG_ON(next->vm_start > end);
2576 #endif
2577         } for_each_vma_range(*vmi, next, end);
2578
2579 #if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2580         /* Make sure no VMAs are about to be lost. */
2581         {
2582                 MA_STATE(test, &mt_detach, 0, 0);
2583                 struct vm_area_struct *vma_mas, *vma_test;
2584                 int test_count = 0;
2585
2586                 vma_iter_set(vmi, start);
2587                 rcu_read_lock();
2588                 vma_test = mas_find(&test, count - 1);
2589                 for_each_vma_range(*vmi, vma_mas, end) {
2590                         BUG_ON(vma_mas != vma_test);
2591                         test_count++;
2592                         vma_test = mas_next(&test, count - 1);
2593                 }
2594                 rcu_read_unlock();
2595                 BUG_ON(count != test_count);
2596         }
2597 #endif
2598
2599         while (vma_iter_addr(vmi) > start)
2600                 vma_iter_prev_range(vmi);
2601
2602         error = vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL);
2603         if (error)
2604                 goto clear_tree_failed;
2605
2606         /* Point of no return */
2607         mm->locked_vm -= locked_vm;
2608         mm->map_count -= count;
2609         if (unlock)
2610                 mmap_write_downgrade(mm);
2611
2612         prev = vma_iter_prev_range(vmi);
2613         next = vma_next(vmi);
2614         if (next)
2615                 vma_iter_prev_range(vmi);
2616
2617         /*
2618          * We can free page tables without write-locking mmap_lock because VMAs
2619          * were isolated before we downgraded mmap_lock.
2620          */
2621         mas_set(&mas_detach, 1);
2622         unmap_region(mm, &mas_detach, vma, prev, next, start, end, count,
2623                      !unlock);
2624         /* Statistics and freeing VMAs */
2625         mas_set(&mas_detach, 0);
2626         remove_mt(mm, &mas_detach);
2627         validate_mm(mm);
2628         if (unlock)
2629                 mmap_read_unlock(mm);
2630
2631         __mt_destroy(&mt_detach);
2632         return 0;
2633
2634 clear_tree_failed:
2635 userfaultfd_error:
2636 munmap_gather_failed:
2637 end_split_failed:
2638         mas_set(&mas_detach, 0);
2639         mas_for_each(&mas_detach, next, end)
2640                 vma_mark_detached(next, false);
2641
2642         __mt_destroy(&mt_detach);
2643 start_split_failed:
2644 map_count_exceeded:
2645         validate_mm(mm);
2646         return error;
2647 }
2648
2649 /*
2650  * do_vmi_munmap() - munmap a given range.
2651  * @vmi: The vma iterator
2652  * @mm: The mm_struct
2653  * @start: The start address to munmap
2654  * @len: The length of the range to munmap
2655  * @uf: The userfaultfd list_head
2656  * @unlock: set to true if the user wants to drop the mmap_lock on success
2657  *
2658  * This function takes a @mas that is either pointing to the previous VMA or set
2659  * to MA_START and sets it up to remove the mapping(s).  The @len will be
2660  * aligned and any arch_unmap work will be preformed.
2661  *
2662  * Return: 0 on success and drops the lock if so directed, error and leaves the
2663  * lock held otherwise.
2664  */
2665 int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
2666                   unsigned long start, size_t len, struct list_head *uf,
2667                   bool unlock)
2668 {
2669         unsigned long end;
2670         struct vm_area_struct *vma;
2671
2672         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2673                 return -EINVAL;
2674
2675         end = start + PAGE_ALIGN(len);
2676         if (end == start)
2677                 return -EINVAL;
2678
2679          /* arch_unmap() might do unmaps itself.  */
2680         arch_unmap(mm, start, end);
2681
2682         /* Find the first overlapping VMA */
2683         vma = vma_find(vmi, end);
2684         if (!vma) {
2685                 if (unlock)
2686                         mmap_write_unlock(mm);
2687                 return 0;
2688         }
2689
2690         return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
2691 }
2692
2693 /* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2694  * @mm: The mm_struct
2695  * @start: The start address to munmap
2696  * @len: The length to be munmapped.
2697  * @uf: The userfaultfd list_head
2698  *
2699  * Return: 0 on success, error otherwise.
2700  */
2701 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2702               struct list_head *uf)
2703 {
2704         VMA_ITERATOR(vmi, mm, start);
2705
2706         return do_vmi_munmap(&vmi, mm, start, len, uf, false);
2707 }
2708
2709 unsigned long mmap_region(struct file *file, unsigned long addr,
2710                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2711                 struct list_head *uf)
2712 {
2713         struct mm_struct *mm = current->mm;
2714         struct vm_area_struct *vma = NULL;
2715         struct vm_area_struct *next, *prev, *merge;
2716         pgoff_t pglen = len >> PAGE_SHIFT;
2717         unsigned long charged = 0;
2718         unsigned long end = addr + len;
2719         unsigned long merge_start = addr, merge_end = end;
2720         bool writable_file_mapping = false;
2721         pgoff_t vm_pgoff;
2722         int error;
2723         VMA_ITERATOR(vmi, mm, addr);
2724
2725         /* Check against address space limit. */
2726         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2727                 unsigned long nr_pages;
2728
2729                 /*
2730                  * MAP_FIXED may remove pages of mappings that intersects with
2731                  * requested mapping. Account for the pages it would unmap.
2732                  */
2733                 nr_pages = count_vma_pages_range(mm, addr, end);
2734
2735                 if (!may_expand_vm(mm, vm_flags,
2736                                         (len >> PAGE_SHIFT) - nr_pages))
2737                         return -ENOMEM;
2738         }
2739
2740         /* Unmap any existing mapping in the area */
2741         if (do_vmi_munmap(&vmi, mm, addr, len, uf, false))
2742                 return -ENOMEM;
2743
2744         /*
2745          * Private writable mapping: check memory availability
2746          */
2747         if (accountable_mapping(file, vm_flags)) {
2748                 charged = len >> PAGE_SHIFT;
2749                 if (security_vm_enough_memory_mm(mm, charged))
2750                         return -ENOMEM;
2751                 vm_flags |= VM_ACCOUNT;
2752         }
2753
2754         next = vma_next(&vmi);
2755         prev = vma_prev(&vmi);
2756         if (vm_flags & VM_SPECIAL) {
2757                 if (prev)
2758                         vma_iter_next_range(&vmi);
2759                 goto cannot_expand;
2760         }
2761
2762         /* Attempt to expand an old mapping */
2763         /* Check next */
2764         if (next && next->vm_start == end && !vma_policy(next) &&
2765             can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2766                                  NULL_VM_UFFD_CTX, NULL)) {
2767                 merge_end = next->vm_end;
2768                 vma = next;
2769                 vm_pgoff = next->vm_pgoff - pglen;
2770         }
2771
2772         /* Check prev */
2773         if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2774             (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2775                                        pgoff, vma->vm_userfaultfd_ctx, NULL) :
2776                    can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2777                                        NULL_VM_UFFD_CTX, NULL))) {
2778                 merge_start = prev->vm_start;
2779                 vma = prev;
2780                 vm_pgoff = prev->vm_pgoff;
2781         } else if (prev) {
2782                 vma_iter_next_range(&vmi);
2783         }
2784
2785         /* Actually expand, if possible */
2786         if (vma &&
2787             !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) {
2788                 khugepaged_enter_vma(vma, vm_flags);
2789                 goto expanded;
2790         }
2791
2792         if (vma == prev)
2793                 vma_iter_set(&vmi, addr);
2794 cannot_expand:
2795
2796         /*
2797          * Determine the object being mapped and call the appropriate
2798          * specific mapper. the address has already been validated, but
2799          * not unmapped, but the maps are removed from the list.
2800          */
2801         vma = vm_area_alloc(mm);
2802         if (!vma) {
2803                 error = -ENOMEM;
2804                 goto unacct_error;
2805         }
2806
2807         vma_iter_config(&vmi, addr, end);
2808         vma->vm_start = addr;
2809         vma->vm_end = end;
2810         vm_flags_init(vma, vm_flags);
2811         vma->vm_page_prot = vm_get_page_prot(vm_flags);
2812         vma->vm_pgoff = pgoff;
2813
2814         if (file) {
2815                 vma->vm_file = get_file(file);
2816                 error = call_mmap(file, vma);
2817                 if (error)
2818                         goto unmap_and_free_vma;
2819
2820                 if (vma_is_shared_maywrite(vma)) {
2821                         error = mapping_map_writable(file->f_mapping);
2822                         if (error)
2823                                 goto close_and_free_vma;
2824
2825                         writable_file_mapping = true;
2826                 }
2827
2828                 /*
2829                  * Expansion is handled above, merging is handled below.
2830                  * Drivers should not alter the address of the VMA.
2831                  */
2832                 error = -EINVAL;
2833                 if (WARN_ON((addr != vma->vm_start)))
2834                         goto close_and_free_vma;
2835
2836                 vma_iter_config(&vmi, addr, end);
2837                 /*
2838                  * If vm_flags changed after call_mmap(), we should try merge
2839                  * vma again as we may succeed this time.
2840                  */
2841                 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2842                         merge = vma_merge_new_vma(&vmi, prev, vma,
2843                                                   vma->vm_start, vma->vm_end,
2844                                                   vma->vm_pgoff);
2845                         if (merge) {
2846                                 /*
2847                                  * ->mmap() can change vma->vm_file and fput
2848                                  * the original file. So fput the vma->vm_file
2849                                  * here or we would add an extra fput for file
2850                                  * and cause general protection fault
2851                                  * ultimately.
2852                                  */
2853                                 fput(vma->vm_file);
2854                                 vm_area_free(vma);
2855                                 vma = merge;
2856                                 /* Update vm_flags to pick up the change. */
2857                                 vm_flags = vma->vm_flags;
2858                                 goto unmap_writable;
2859                         }
2860                 }
2861
2862                 vm_flags = vma->vm_flags;
2863         } else if (vm_flags & VM_SHARED) {
2864                 error = shmem_zero_setup(vma);
2865                 if (error)
2866                         goto free_vma;
2867         } else {
2868                 vma_set_anonymous(vma);
2869         }
2870
2871         if (map_deny_write_exec(vma, vma->vm_flags)) {
2872                 error = -EACCES;
2873                 goto close_and_free_vma;
2874         }
2875
2876         /* Allow architectures to sanity-check the vm_flags */
2877         error = -EINVAL;
2878         if (!arch_validate_flags(vma->vm_flags))
2879                 goto close_and_free_vma;
2880
2881         error = -ENOMEM;
2882         if (vma_iter_prealloc(&vmi, vma))
2883                 goto close_and_free_vma;
2884
2885         /* Lock the VMA since it is modified after insertion into VMA tree */
2886         vma_start_write(vma);
2887         vma_iter_store(&vmi, vma);
2888         mm->map_count++;
2889         if (vma->vm_file) {
2890                 i_mmap_lock_write(vma->vm_file->f_mapping);
2891                 if (vma_is_shared_maywrite(vma))
2892                         mapping_allow_writable(vma->vm_file->f_mapping);
2893
2894                 flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2895                 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2896                 flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2897                 i_mmap_unlock_write(vma->vm_file->f_mapping);
2898         }
2899
2900         /*
2901          * vma_merge() calls khugepaged_enter_vma() either, the below
2902          * call covers the non-merge case.
2903          */
2904         khugepaged_enter_vma(vma, vma->vm_flags);
2905
2906         /* Once vma denies write, undo our temporary denial count */
2907 unmap_writable:
2908         if (writable_file_mapping)
2909                 mapping_unmap_writable(file->f_mapping);
2910         file = vma->vm_file;
2911         ksm_add_vma(vma);
2912 expanded:
2913         perf_event_mmap(vma);
2914
2915         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2916         if (vm_flags & VM_LOCKED) {
2917                 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2918                                         is_vm_hugetlb_page(vma) ||
2919                                         vma == get_gate_vma(current->mm))
2920                         vm_flags_clear(vma, VM_LOCKED_MASK);
2921                 else
2922                         mm->locked_vm += (len >> PAGE_SHIFT);
2923         }
2924
2925         if (file)
2926                 uprobe_mmap(vma);
2927
2928         /*
2929          * New (or expanded) vma always get soft dirty status.
2930          * Otherwise user-space soft-dirty page tracker won't
2931          * be able to distinguish situation when vma area unmapped,
2932          * then new mapped in-place (which must be aimed as
2933          * a completely new data area).
2934          */
2935         vm_flags_set(vma, VM_SOFTDIRTY);
2936
2937         vma_set_page_prot(vma);
2938
2939         validate_mm(mm);
2940         return addr;
2941
2942 close_and_free_vma:
2943         if (file && vma->vm_ops && vma->vm_ops->close)
2944                 vma->vm_ops->close(vma);
2945
2946         if (file || vma->vm_file) {
2947 unmap_and_free_vma:
2948                 fput(vma->vm_file);
2949                 vma->vm_file = NULL;
2950
2951                 vma_iter_set(&vmi, vma->vm_end);
2952                 /* Undo any partial mapping done by a device driver. */
2953                 unmap_region(mm, &vmi.mas, vma, prev, next, vma->vm_start,
2954                              vma->vm_end, vma->vm_end, true);
2955         }
2956         if (writable_file_mapping)
2957                 mapping_unmap_writable(file->f_mapping);
2958 free_vma:
2959         vm_area_free(vma);
2960 unacct_error:
2961         if (charged)
2962                 vm_unacct_memory(charged);
2963         validate_mm(mm);
2964         return error;
2965 }
2966
2967 static int __vm_munmap(unsigned long start, size_t len, bool unlock)
2968 {
2969         int ret;
2970         struct mm_struct *mm = current->mm;
2971         LIST_HEAD(uf);
2972         VMA_ITERATOR(vmi, mm, start);
2973
2974         if (mmap_write_lock_killable(mm))
2975                 return -EINTR;
2976
2977         ret = do_vmi_munmap(&vmi, mm, start, len, &uf, unlock);
2978         if (ret || !unlock)
2979                 mmap_write_unlock(mm);
2980
2981         userfaultfd_unmap_complete(mm, &uf);
2982         return ret;
2983 }
2984
2985 int vm_munmap(unsigned long start, size_t len)
2986 {
2987         return __vm_munmap(start, len, false);
2988 }
2989 EXPORT_SYMBOL(vm_munmap);
2990
2991 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2992 {
2993         addr = untagged_addr(addr);
2994         return __vm_munmap(addr, len, true);
2995 }
2996
2997
2998 /*
2999  * Emulation of deprecated remap_file_pages() syscall.
3000  */
3001 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
3002                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
3003 {
3004
3005         struct mm_struct *mm = current->mm;
3006         struct vm_area_struct *vma;
3007         unsigned long populate = 0;
3008         unsigned long ret = -EINVAL;
3009         struct file *file;
3010
3011         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
3012                      current->comm, current->pid);
3013
3014         if (prot)
3015                 return ret;
3016         start = start & PAGE_MASK;
3017         size = size & PAGE_MASK;
3018
3019         if (start + size <= start)
3020                 return ret;
3021
3022         /* Does pgoff wrap? */
3023         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
3024                 return ret;
3025
3026         if (mmap_write_lock_killable(mm))
3027                 return -EINTR;
3028
3029         vma = vma_lookup(mm, start);
3030
3031         if (!vma || !(vma->vm_flags & VM_SHARED))
3032                 goto out;
3033
3034         if (start + size > vma->vm_end) {
3035                 VMA_ITERATOR(vmi, mm, vma->vm_end);
3036                 struct vm_area_struct *next, *prev = vma;
3037
3038                 for_each_vma_range(vmi, next, start + size) {
3039                         /* hole between vmas ? */
3040                         if (next->vm_start != prev->vm_end)
3041                                 goto out;
3042
3043                         if (next->vm_file != vma->vm_file)
3044                                 goto out;
3045
3046                         if (next->vm_flags != vma->vm_flags)
3047                                 goto out;
3048
3049                         if (start + size <= next->vm_end)
3050                                 break;
3051
3052                         prev = next;
3053                 }
3054
3055                 if (!next)
3056                         goto out;
3057         }
3058
3059         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3060         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3061         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3062
3063         flags &= MAP_NONBLOCK;
3064         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3065         if (vma->vm_flags & VM_LOCKED)
3066                 flags |= MAP_LOCKED;
3067
3068         file = get_file(vma->vm_file);
3069         ret = do_mmap(vma->vm_file, start, size,
3070                         prot, flags, 0, pgoff, &populate, NULL);
3071         fput(file);
3072 out:
3073         mmap_write_unlock(mm);
3074         if (populate)
3075                 mm_populate(ret, populate);
3076         if (!IS_ERR_VALUE(ret))
3077                 ret = 0;
3078         return ret;
3079 }
3080
3081 /*
3082  * do_vma_munmap() - Unmap a full or partial vma.
3083  * @vmi: The vma iterator pointing at the vma
3084  * @vma: The first vma to be munmapped
3085  * @start: the start of the address to unmap
3086  * @end: The end of the address to unmap
3087  * @uf: The userfaultfd list_head
3088  * @unlock: Drop the lock on success
3089  *
3090  * unmaps a VMA mapping when the vma iterator is already in position.
3091  * Does not handle alignment.
3092  *
3093  * Return: 0 on success drops the lock of so directed, error on failure and will
3094  * still hold the lock.
3095  */
3096 int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
3097                 unsigned long start, unsigned long end, struct list_head *uf,
3098                 bool unlock)
3099 {
3100         struct mm_struct *mm = vma->vm_mm;
3101
3102         arch_unmap(mm, start, end);
3103         return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
3104 }
3105
3106 /*
3107  * do_brk_flags() - Increase the brk vma if the flags match.
3108  * @vmi: The vma iterator
3109  * @addr: The start address
3110  * @len: The length of the increase
3111  * @vma: The vma,
3112  * @flags: The VMA Flags
3113  *
3114  * Extend the brk VMA from addr to addr + len.  If the VMA is NULL or the flags
3115  * do not match then create a new anonymous VMA.  Eventually we may be able to
3116  * do some brk-specific accounting here.
3117  */
3118 static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
3119                 unsigned long addr, unsigned long len, unsigned long flags)
3120 {
3121         struct mm_struct *mm = current->mm;
3122         struct vma_prepare vp;
3123
3124         /*
3125          * Check against address space limits by the changed size
3126          * Note: This happens *after* clearing old mappings in some code paths.
3127          */
3128         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3129         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3130                 return -ENOMEM;
3131
3132         if (mm->map_count > sysctl_max_map_count)
3133                 return -ENOMEM;
3134
3135         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3136                 return -ENOMEM;
3137
3138         /*
3139          * Expand the existing vma if possible; Note that singular lists do not
3140          * occur after forking, so the expand will only happen on new VMAs.
3141          */
3142         if (vma && vma->vm_end == addr && !vma_policy(vma) &&
3143             can_vma_merge_after(vma, flags, NULL, NULL,
3144                                 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
3145                 vma_iter_config(vmi, vma->vm_start, addr + len);
3146                 if (vma_iter_prealloc(vmi, vma))
3147                         goto unacct_fail;
3148
3149                 vma_start_write(vma);
3150
3151                 init_vma_prep(&vp, vma);
3152                 vma_prepare(&vp);
3153                 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
3154                 vma->vm_end = addr + len;
3155                 vm_flags_set(vma, VM_SOFTDIRTY);
3156                 vma_iter_store(vmi, vma);
3157
3158                 vma_complete(&vp, vmi, mm);
3159                 khugepaged_enter_vma(vma, flags);
3160                 goto out;
3161         }
3162
3163         if (vma)
3164                 vma_iter_next_range(vmi);
3165         /* create a vma struct for an anonymous mapping */
3166         vma = vm_area_alloc(mm);
3167         if (!vma)
3168                 goto unacct_fail;
3169
3170         vma_set_anonymous(vma);
3171         vma->vm_start = addr;
3172         vma->vm_end = addr + len;
3173         vma->vm_pgoff = addr >> PAGE_SHIFT;
3174         vm_flags_init(vma, flags);
3175         vma->vm_page_prot = vm_get_page_prot(flags);
3176         vma_start_write(vma);
3177         if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
3178                 goto mas_store_fail;
3179
3180         mm->map_count++;
3181         validate_mm(mm);
3182         ksm_add_vma(vma);
3183 out:
3184         perf_event_mmap(vma);
3185         mm->total_vm += len >> PAGE_SHIFT;
3186         mm->data_vm += len >> PAGE_SHIFT;
3187         if (flags & VM_LOCKED)
3188                 mm->locked_vm += (len >> PAGE_SHIFT);
3189         vm_flags_set(vma, VM_SOFTDIRTY);
3190         return 0;
3191
3192 mas_store_fail:
3193         vm_area_free(vma);
3194 unacct_fail:
3195         vm_unacct_memory(len >> PAGE_SHIFT);
3196         return -ENOMEM;
3197 }
3198
3199 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3200 {
3201         struct mm_struct *mm = current->mm;
3202         struct vm_area_struct *vma = NULL;
3203         unsigned long len;
3204         int ret;
3205         bool populate;
3206         LIST_HEAD(uf);
3207         VMA_ITERATOR(vmi, mm, addr);
3208
3209         len = PAGE_ALIGN(request);
3210         if (len < request)
3211                 return -ENOMEM;
3212         if (!len)
3213                 return 0;
3214
3215         /* Until we need other flags, refuse anything except VM_EXEC. */
3216         if ((flags & (~VM_EXEC)) != 0)
3217                 return -EINVAL;
3218
3219         if (mmap_write_lock_killable(mm))
3220                 return -EINTR;
3221
3222         ret = check_brk_limits(addr, len);
3223         if (ret)
3224                 goto limits_failed;
3225
3226         ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
3227         if (ret)
3228                 goto munmap_failed;
3229
3230         vma = vma_prev(&vmi);
3231         ret = do_brk_flags(&vmi, vma, addr, len, flags);
3232         populate = ((mm->def_flags & VM_LOCKED) != 0);
3233         mmap_write_unlock(mm);
3234         userfaultfd_unmap_complete(mm, &uf);
3235         if (populate && !ret)
3236                 mm_populate(addr, len);
3237         return ret;
3238
3239 munmap_failed:
3240 limits_failed:
3241         mmap_write_unlock(mm);
3242         return ret;
3243 }
3244 EXPORT_SYMBOL(vm_brk_flags);
3245
3246 /* Release all mmaps. */
3247 void exit_mmap(struct mm_struct *mm)
3248 {
3249         struct mmu_gather tlb;
3250         struct vm_area_struct *vma;
3251         unsigned long nr_accounted = 0;
3252         MA_STATE(mas, &mm->mm_mt, 0, 0);
3253         int count = 0;
3254
3255         /* mm's last user has gone, and its about to be pulled down */
3256         mmu_notifier_release(mm);
3257
3258         mmap_read_lock(mm);
3259         arch_exit_mmap(mm);
3260
3261         vma = mas_find(&mas, ULONG_MAX);
3262         if (!vma || unlikely(xa_is_zero(vma))) {
3263                 /* Can happen if dup_mmap() received an OOM */
3264                 mmap_read_unlock(mm);
3265                 mmap_write_lock(mm);
3266                 goto destroy;
3267         }
3268
3269         lru_add_drain();
3270         flush_cache_mm(mm);
3271         tlb_gather_mmu_fullmm(&tlb, mm);
3272         /* update_hiwater_rss(mm) here? but nobody should be looking */
3273         /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3274         unmap_vmas(&tlb, &mas, vma, 0, ULONG_MAX, ULONG_MAX, false);
3275         mmap_read_unlock(mm);
3276
3277         /*
3278          * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3279          * because the memory has been already freed.
3280          */
3281         set_bit(MMF_OOM_SKIP, &mm->flags);
3282         mmap_write_lock(mm);
3283         mt_clear_in_rcu(&mm->mm_mt);
3284         mas_set(&mas, vma->vm_end);
3285         free_pgtables(&tlb, &mas, vma, FIRST_USER_ADDRESS,
3286                       USER_PGTABLES_CEILING, true);
3287         tlb_finish_mmu(&tlb);
3288
3289         /*
3290          * Walk the list again, actually closing and freeing it, with preemption
3291          * enabled, without holding any MM locks besides the unreachable
3292          * mmap_write_lock.
3293          */
3294         mas_set(&mas, vma->vm_end);
3295         do {
3296                 if (vma->vm_flags & VM_ACCOUNT)
3297                         nr_accounted += vma_pages(vma);
3298                 remove_vma(vma, true);
3299                 count++;
3300                 cond_resched();
3301                 vma = mas_find(&mas, ULONG_MAX);
3302         } while (vma && likely(!xa_is_zero(vma)));
3303
3304         BUG_ON(count != mm->map_count);
3305
3306         trace_exit_mmap(mm);
3307 destroy:
3308         __mt_destroy(&mm->mm_mt);
3309         mmap_write_unlock(mm);
3310         vm_unacct_memory(nr_accounted);
3311 }
3312
3313 /* Insert vm structure into process list sorted by address
3314  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3315  * then i_mmap_rwsem is taken here.
3316  */
3317 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3318 {
3319         unsigned long charged = vma_pages(vma);
3320
3321
3322         if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3323                 return -ENOMEM;
3324
3325         if ((vma->vm_flags & VM_ACCOUNT) &&
3326              security_vm_enough_memory_mm(mm, charged))
3327                 return -ENOMEM;
3328
3329         /*
3330          * The vm_pgoff of a purely anonymous vma should be irrelevant
3331          * until its first write fault, when page's anon_vma and index
3332          * are set.  But now set the vm_pgoff it will almost certainly
3333          * end up with (unless mremap moves it elsewhere before that
3334          * first wfault), so /proc/pid/maps tells a consistent story.
3335          *
3336          * By setting it to reflect the virtual start address of the
3337          * vma, merges and splits can happen in a seamless way, just
3338          * using the existing file pgoff checks and manipulations.
3339          * Similarly in do_mmap and in do_brk_flags.
3340          */
3341         if (vma_is_anonymous(vma)) {
3342                 BUG_ON(vma->anon_vma);
3343                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3344         }
3345
3346         if (vma_link(mm, vma)) {
3347                 if (vma->vm_flags & VM_ACCOUNT)
3348                         vm_unacct_memory(charged);
3349                 return -ENOMEM;
3350         }
3351
3352         return 0;
3353 }
3354
3355 /*
3356  * Copy the vma structure to a new location in the same mm,
3357  * prior to moving page table entries, to effect an mremap move.
3358  */
3359 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3360         unsigned long addr, unsigned long len, pgoff_t pgoff,
3361         bool *need_rmap_locks)
3362 {
3363         struct vm_area_struct *vma = *vmap;
3364         unsigned long vma_start = vma->vm_start;
3365         struct mm_struct *mm = vma->vm_mm;
3366         struct vm_area_struct *new_vma, *prev;
3367         bool faulted_in_anon_vma = true;
3368         VMA_ITERATOR(vmi, mm, addr);
3369
3370         /*
3371          * If anonymous vma has not yet been faulted, update new pgoff
3372          * to match new location, to increase its chance of merging.
3373          */
3374         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3375                 pgoff = addr >> PAGE_SHIFT;
3376                 faulted_in_anon_vma = false;
3377         }
3378
3379         new_vma = find_vma_prev(mm, addr, &prev);
3380         if (new_vma && new_vma->vm_start < addr + len)
3381                 return NULL;    /* should never get here */
3382
3383         new_vma = vma_merge_new_vma(&vmi, prev, vma, addr, addr + len, pgoff);
3384         if (new_vma) {
3385                 /*
3386                  * Source vma may have been merged into new_vma
3387                  */
3388                 if (unlikely(vma_start >= new_vma->vm_start &&
3389                              vma_start < new_vma->vm_end)) {
3390                         /*
3391                          * The only way we can get a vma_merge with
3392                          * self during an mremap is if the vma hasn't
3393                          * been faulted in yet and we were allowed to
3394                          * reset the dst vma->vm_pgoff to the
3395                          * destination address of the mremap to allow
3396                          * the merge to happen. mremap must change the
3397                          * vm_pgoff linearity between src and dst vmas
3398                          * (in turn preventing a vma_merge) to be
3399                          * safe. It is only safe to keep the vm_pgoff
3400                          * linear if there are no pages mapped yet.
3401                          */
3402                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3403                         *vmap = vma = new_vma;
3404                 }
3405                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3406         } else {
3407                 new_vma = vm_area_dup(vma);
3408                 if (!new_vma)
3409                         goto out;
3410                 new_vma->vm_start = addr;
3411                 new_vma->vm_end = addr + len;
3412                 new_vma->vm_pgoff = pgoff;
3413                 if (vma_dup_policy(vma, new_vma))
3414                         goto out_free_vma;
3415                 if (anon_vma_clone(new_vma, vma))
3416                         goto out_free_mempol;
3417                 if (new_vma->vm_file)
3418                         get_file(new_vma->vm_file);
3419                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3420                         new_vma->vm_ops->open(new_vma);
3421                 if (vma_link(mm, new_vma))
3422                         goto out_vma_link;
3423                 *need_rmap_locks = false;
3424         }
3425         return new_vma;
3426
3427 out_vma_link:
3428         if (new_vma->vm_ops && new_vma->vm_ops->close)
3429                 new_vma->vm_ops->close(new_vma);
3430
3431         if (new_vma->vm_file)
3432                 fput(new_vma->vm_file);
3433
3434         unlink_anon_vmas(new_vma);
3435 out_free_mempol:
3436         mpol_put(vma_policy(new_vma));
3437 out_free_vma:
3438         vm_area_free(new_vma);
3439 out:
3440         return NULL;
3441 }
3442
3443 /*
3444  * Return true if the calling process may expand its vm space by the passed
3445  * number of pages
3446  */
3447 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3448 {
3449         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3450                 return false;
3451
3452         if (is_data_mapping(flags) &&
3453             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3454                 /* Workaround for Valgrind */
3455                 if (rlimit(RLIMIT_DATA) == 0 &&
3456                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3457                         return true;
3458
3459                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3460                              current->comm, current->pid,
3461                              (mm->data_vm + npages) << PAGE_SHIFT,
3462                              rlimit(RLIMIT_DATA),
3463                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3464
3465                 if (!ignore_rlimit_data)
3466                         return false;
3467         }
3468
3469         return true;
3470 }
3471
3472 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3473 {
3474         WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3475
3476         if (is_exec_mapping(flags))
3477                 mm->exec_vm += npages;
3478         else if (is_stack_mapping(flags))
3479                 mm->stack_vm += npages;
3480         else if (is_data_mapping(flags))
3481                 mm->data_vm += npages;
3482 }
3483
3484 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3485
3486 /*
3487  * Having a close hook prevents vma merging regardless of flags.
3488  */
3489 static void special_mapping_close(struct vm_area_struct *vma)
3490 {
3491 }
3492
3493 static const char *special_mapping_name(struct vm_area_struct *vma)
3494 {
3495         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3496 }
3497
3498 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3499 {
3500         struct vm_special_mapping *sm = new_vma->vm_private_data;
3501
3502         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3503                 return -EFAULT;
3504
3505         if (sm->mremap)
3506                 return sm->mremap(sm, new_vma);
3507
3508         return 0;
3509 }
3510
3511 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3512 {
3513         /*
3514          * Forbid splitting special mappings - kernel has expectations over
3515          * the number of pages in mapping. Together with VM_DONTEXPAND
3516          * the size of vma should stay the same over the special mapping's
3517          * lifetime.
3518          */
3519         return -EINVAL;
3520 }
3521
3522 static const struct vm_operations_struct special_mapping_vmops = {
3523         .close = special_mapping_close,
3524         .fault = special_mapping_fault,
3525         .mremap = special_mapping_mremap,
3526         .name = special_mapping_name,
3527         /* vDSO code relies that VVAR can't be accessed remotely */
3528         .access = NULL,
3529         .may_split = special_mapping_split,
3530 };
3531
3532 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3533         .close = special_mapping_close,
3534         .fault = special_mapping_fault,
3535 };
3536
3537 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3538 {
3539         struct vm_area_struct *vma = vmf->vma;
3540         pgoff_t pgoff;
3541         struct page **pages;
3542
3543         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3544                 pages = vma->vm_private_data;
3545         } else {
3546                 struct vm_special_mapping *sm = vma->vm_private_data;
3547
3548                 if (sm->fault)
3549                         return sm->fault(sm, vmf->vma, vmf);
3550
3551                 pages = sm->pages;
3552         }
3553
3554         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3555                 pgoff--;
3556
3557         if (*pages) {
3558                 struct page *page = *pages;
3559                 get_page(page);
3560                 vmf->page = page;
3561                 return 0;
3562         }
3563
3564         return VM_FAULT_SIGBUS;
3565 }
3566
3567 static struct vm_area_struct *__install_special_mapping(
3568         struct mm_struct *mm,
3569         unsigned long addr, unsigned long len,
3570         unsigned long vm_flags, void *priv,
3571         const struct vm_operations_struct *ops)
3572 {
3573         int ret;
3574         struct vm_area_struct *vma;
3575
3576         vma = vm_area_alloc(mm);
3577         if (unlikely(vma == NULL))
3578                 return ERR_PTR(-ENOMEM);
3579
3580         vma->vm_start = addr;
3581         vma->vm_end = addr + len;
3582
3583         vm_flags_init(vma, (vm_flags | mm->def_flags |
3584                       VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
3585         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3586
3587         vma->vm_ops = ops;
3588         vma->vm_private_data = priv;
3589
3590         ret = insert_vm_struct(mm, vma);
3591         if (ret)
3592                 goto out;
3593
3594         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3595
3596         perf_event_mmap(vma);
3597
3598         return vma;
3599
3600 out:
3601         vm_area_free(vma);
3602         return ERR_PTR(ret);
3603 }
3604
3605 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3606         const struct vm_special_mapping *sm)
3607 {
3608         return vma->vm_private_data == sm &&
3609                 (vma->vm_ops == &special_mapping_vmops ||
3610                  vma->vm_ops == &legacy_special_mapping_vmops);
3611 }
3612
3613 /*
3614  * Called with mm->mmap_lock held for writing.
3615  * Insert a new vma covering the given region, with the given flags.
3616  * Its pages are supplied by the given array of struct page *.
3617  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3618  * The region past the last page supplied will always produce SIGBUS.
3619  * The array pointer and the pages it points to are assumed to stay alive
3620  * for as long as this mapping might exist.
3621  */
3622 struct vm_area_struct *_install_special_mapping(
3623         struct mm_struct *mm,
3624         unsigned long addr, unsigned long len,
3625         unsigned long vm_flags, const struct vm_special_mapping *spec)
3626 {
3627         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3628                                         &special_mapping_vmops);
3629 }
3630
3631 int install_special_mapping(struct mm_struct *mm,
3632                             unsigned long addr, unsigned long len,
3633                             unsigned long vm_flags, struct page **pages)
3634 {
3635         struct vm_area_struct *vma = __install_special_mapping(
3636                 mm, addr, len, vm_flags, (void *)pages,
3637                 &legacy_special_mapping_vmops);
3638
3639         return PTR_ERR_OR_ZERO(vma);
3640 }
3641
3642 static DEFINE_MUTEX(mm_all_locks_mutex);
3643
3644 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3645 {
3646         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3647                 /*
3648                  * The LSB of head.next can't change from under us
3649                  * because we hold the mm_all_locks_mutex.
3650                  */
3651                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3652                 /*
3653                  * We can safely modify head.next after taking the
3654                  * anon_vma->root->rwsem. If some other vma in this mm shares
3655                  * the same anon_vma we won't take it again.
3656                  *
3657                  * No need of atomic instructions here, head.next
3658                  * can't change from under us thanks to the
3659                  * anon_vma->root->rwsem.
3660                  */
3661                 if (__test_and_set_bit(0, (unsigned long *)
3662                                        &anon_vma->root->rb_root.rb_root.rb_node))
3663                         BUG();
3664         }
3665 }
3666
3667 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3668 {
3669         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3670                 /*
3671                  * AS_MM_ALL_LOCKS can't change from under us because
3672                  * we hold the mm_all_locks_mutex.
3673                  *
3674                  * Operations on ->flags have to be atomic because
3675                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3676                  * mm_all_locks_mutex, there may be other cpus
3677                  * changing other bitflags in parallel to us.
3678                  */
3679                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3680                         BUG();
3681                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3682         }
3683 }
3684
3685 /*
3686  * This operation locks against the VM for all pte/vma/mm related
3687  * operations that could ever happen on a certain mm. This includes
3688  * vmtruncate, try_to_unmap, and all page faults.
3689  *
3690  * The caller must take the mmap_lock in write mode before calling
3691  * mm_take_all_locks(). The caller isn't allowed to release the
3692  * mmap_lock until mm_drop_all_locks() returns.
3693  *
3694  * mmap_lock in write mode is required in order to block all operations
3695  * that could modify pagetables and free pages without need of
3696  * altering the vma layout. It's also needed in write mode to avoid new
3697  * anon_vmas to be associated with existing vmas.
3698  *
3699  * A single task can't take more than one mm_take_all_locks() in a row
3700  * or it would deadlock.
3701  *
3702  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3703  * mapping->flags avoid to take the same lock twice, if more than one
3704  * vma in this mm is backed by the same anon_vma or address_space.
3705  *
3706  * We take locks in following order, accordingly to comment at beginning
3707  * of mm/rmap.c:
3708  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3709  *     hugetlb mapping);
3710  *   - all vmas marked locked
3711  *   - all i_mmap_rwsem locks;
3712  *   - all anon_vma->rwseml
3713  *
3714  * We can take all locks within these types randomly because the VM code
3715  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3716  * mm_all_locks_mutex.
3717  *
3718  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3719  * that may have to take thousand of locks.
3720  *
3721  * mm_take_all_locks() can fail if it's interrupted by signals.
3722  */
3723 int mm_take_all_locks(struct mm_struct *mm)
3724 {
3725         struct vm_area_struct *vma;
3726         struct anon_vma_chain *avc;
3727         MA_STATE(mas, &mm->mm_mt, 0, 0);
3728
3729         mmap_assert_write_locked(mm);
3730
3731         mutex_lock(&mm_all_locks_mutex);
3732
3733         /*
3734          * vma_start_write() does not have a complement in mm_drop_all_locks()
3735          * because vma_start_write() is always asymmetrical; it marks a VMA as
3736          * being written to until mmap_write_unlock() or mmap_write_downgrade()
3737          * is reached.
3738          */
3739         mas_for_each(&mas, vma, ULONG_MAX) {
3740                 if (signal_pending(current))
3741                         goto out_unlock;
3742                 vma_start_write(vma);
3743         }
3744
3745         mas_set(&mas, 0);
3746         mas_for_each(&mas, vma, ULONG_MAX) {
3747                 if (signal_pending(current))
3748                         goto out_unlock;
3749                 if (vma->vm_file && vma->vm_file->f_mapping &&
3750                                 is_vm_hugetlb_page(vma))
3751                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3752         }
3753
3754         mas_set(&mas, 0);
3755         mas_for_each(&mas, vma, ULONG_MAX) {
3756                 if (signal_pending(current))
3757                         goto out_unlock;
3758                 if (vma->vm_file && vma->vm_file->f_mapping &&
3759                                 !is_vm_hugetlb_page(vma))
3760                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3761         }
3762
3763         mas_set(&mas, 0);
3764         mas_for_each(&mas, vma, ULONG_MAX) {
3765                 if (signal_pending(current))
3766                         goto out_unlock;
3767                 if (vma->anon_vma)
3768                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3769                                 vm_lock_anon_vma(mm, avc->anon_vma);
3770         }
3771
3772         return 0;
3773
3774 out_unlock:
3775         mm_drop_all_locks(mm);
3776         return -EINTR;
3777 }
3778
3779 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3780 {
3781         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3782                 /*
3783                  * The LSB of head.next can't change to 0 from under
3784                  * us because we hold the mm_all_locks_mutex.
3785                  *
3786                  * We must however clear the bitflag before unlocking
3787                  * the vma so the users using the anon_vma->rb_root will
3788                  * never see our bitflag.
3789                  *
3790                  * No need of atomic instructions here, head.next
3791                  * can't change from under us until we release the
3792                  * anon_vma->root->rwsem.
3793                  */
3794                 if (!__test_and_clear_bit(0, (unsigned long *)
3795                                           &anon_vma->root->rb_root.rb_root.rb_node))
3796                         BUG();
3797                 anon_vma_unlock_write(anon_vma);
3798         }
3799 }
3800
3801 static void vm_unlock_mapping(struct address_space *mapping)
3802 {
3803         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3804                 /*
3805                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3806                  * because we hold the mm_all_locks_mutex.
3807                  */
3808                 i_mmap_unlock_write(mapping);
3809                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3810                                         &mapping->flags))
3811                         BUG();
3812         }
3813 }
3814
3815 /*
3816  * The mmap_lock cannot be released by the caller until
3817  * mm_drop_all_locks() returns.
3818  */
3819 void mm_drop_all_locks(struct mm_struct *mm)
3820 {
3821         struct vm_area_struct *vma;
3822         struct anon_vma_chain *avc;
3823         MA_STATE(mas, &mm->mm_mt, 0, 0);
3824
3825         mmap_assert_write_locked(mm);
3826         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3827
3828         mas_for_each(&mas, vma, ULONG_MAX) {
3829                 if (vma->anon_vma)
3830                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3831                                 vm_unlock_anon_vma(avc->anon_vma);
3832                 if (vma->vm_file && vma->vm_file->f_mapping)
3833                         vm_unlock_mapping(vma->vm_file->f_mapping);
3834         }
3835
3836         mutex_unlock(&mm_all_locks_mutex);
3837 }
3838
3839 /*
3840  * initialise the percpu counter for VM
3841  */
3842 void __init mmap_init(void)
3843 {
3844         int ret;
3845
3846         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3847         VM_BUG_ON(ret);
3848 }
3849
3850 /*
3851  * Initialise sysctl_user_reserve_kbytes.
3852  *
3853  * This is intended to prevent a user from starting a single memory hogging
3854  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3855  * mode.
3856  *
3857  * The default value is min(3% of free memory, 128MB)
3858  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3859  */
3860 static int init_user_reserve(void)
3861 {
3862         unsigned long free_kbytes;
3863
3864         free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3865
3866         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3867         return 0;
3868 }
3869 subsys_initcall(init_user_reserve);
3870
3871 /*
3872  * Initialise sysctl_admin_reserve_kbytes.
3873  *
3874  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3875  * to log in and kill a memory hogging process.
3876  *
3877  * Systems with more than 256MB will reserve 8MB, enough to recover
3878  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3879  * only reserve 3% of free pages by default.
3880  */
3881 static int init_admin_reserve(void)
3882 {
3883         unsigned long free_kbytes;
3884
3885         free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3886
3887         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3888         return 0;
3889 }
3890 subsys_initcall(init_admin_reserve);
3891
3892 /*
3893  * Reinititalise user and admin reserves if memory is added or removed.
3894  *
3895  * The default user reserve max is 128MB, and the default max for the
3896  * admin reserve is 8MB. These are usually, but not always, enough to
3897  * enable recovery from a memory hogging process using login/sshd, a shell,
3898  * and tools like top. It may make sense to increase or even disable the
3899  * reserve depending on the existence of swap or variations in the recovery
3900  * tools. So, the admin may have changed them.
3901  *
3902  * If memory is added and the reserves have been eliminated or increased above
3903  * the default max, then we'll trust the admin.
3904  *
3905  * If memory is removed and there isn't enough free memory, then we
3906  * need to reset the reserves.
3907  *
3908  * Otherwise keep the reserve set by the admin.
3909  */
3910 static int reserve_mem_notifier(struct notifier_block *nb,
3911                              unsigned long action, void *data)
3912 {
3913         unsigned long tmp, free_kbytes;
3914
3915         switch (action) {
3916         case MEM_ONLINE:
3917                 /* Default max is 128MB. Leave alone if modified by operator. */
3918                 tmp = sysctl_user_reserve_kbytes;
3919                 if (0 < tmp && tmp < (1UL << 17))
3920                         init_user_reserve();
3921
3922                 /* Default max is 8MB.  Leave alone if modified by operator. */
3923                 tmp = sysctl_admin_reserve_kbytes;
3924                 if (0 < tmp && tmp < (1UL << 13))
3925                         init_admin_reserve();
3926
3927                 break;
3928         case MEM_OFFLINE:
3929                 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3930
3931                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3932                         init_user_reserve();
3933                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3934                                 sysctl_user_reserve_kbytes);
3935                 }
3936
3937                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3938                         init_admin_reserve();
3939                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3940                                 sysctl_admin_reserve_kbytes);
3941                 }
3942                 break;
3943         default:
3944                 break;
3945         }
3946         return NOTIFY_OK;
3947 }
3948
3949 static int __meminit init_reserve_notifier(void)
3950 {
3951         if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3952                 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3953
3954         return 0;
3955 }
3956 subsys_initcall(init_reserve_notifier);