Loading...
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
66const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
67const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
69#endif
70#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
74#endif
75
76static bool ignore_rlimit_data;
77core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
78
79static 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
84static 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. */
90void 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 */
107static 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 */
122void 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 */
137static 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
151static 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 */
165static 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}
176static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
177 unsigned long addr, unsigned long request, unsigned long flags);
178SYSCALL_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
266success:
267 mmap_write_unlock(mm);
268success_unlocked:
269 userfaultfd_unmap_complete(mm, &uf);
270 if (populate)
271 mm_populate(oldbrk, newbrk - oldbrk);
272 return brk;
273
274out:
275 mm->brk = origbrk;
276 mmap_write_unlock(mm);
277 return origbrk;
278}
279
280#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
281static 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 */
349static inline void
350anon_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
358static inline void
359anon_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
367static 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
384static 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
395static 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 */
428static 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 */
452static 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 */
463static 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 */
510static 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) {
552again:
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 */
590static 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 */
630int 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
671nomem:
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 */
686int 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 */
720static 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
746static 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 */
772static bool
773can_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 */
795static bool
796can_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 */
863static 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 if (end == curr->vm_end) { /* case 7 */
958 /*
959 * can_vma_merge_after() assumed we would not be
960 * removing prev vma, so it skipped the check
961 * for vm_ops->close, but we are removing curr
962 */
963 if (curr->vm_ops && curr->vm_ops->close)
964 err = -EINVAL;
965 remove = curr;
966 } else { /* case 5 */
967 adjust = curr;
968 adj_start = (end - curr->vm_start);
969 }
970 if (!err)
971 err = dup_anon_vma(prev, curr, &anon_dup);
972 }
973 } else { /* merge_next */
974 vma_start_write(next);
975 res = next;
976 if (prev && addr < prev->vm_end) { /* case 4 */
977 vma_start_write(prev);
978 vma_end = addr;
979 adjust = next;
980 adj_start = -(prev->vm_end - addr);
981 err = dup_anon_vma(next, prev, &anon_dup);
982 } else {
983 /*
984 * Note that cases 3 and 8 are the ONLY ones where prev
985 * is permitted to be (but is not necessarily) NULL.
986 */
987 vma = next; /* case 3 */
988 vma_start = addr;
989 vma_end = next->vm_end;
990 vma_pgoff = next->vm_pgoff - pglen;
991 if (curr) { /* case 8 */
992 vma_pgoff = curr->vm_pgoff;
993 vma_start_write(curr);
994 remove = curr;
995 err = dup_anon_vma(next, curr, &anon_dup);
996 }
997 }
998 }
999
1000 /* Error in anon_vma clone. */
1001 if (err)
1002 goto anon_vma_fail;
1003
1004 if (vma_start < vma->vm_start || vma_end > vma->vm_end)
1005 vma_expanded = true;
1006
1007 if (vma_expanded) {
1008 vma_iter_config(vmi, vma_start, vma_end);
1009 } else {
1010 vma_iter_config(vmi, adjust->vm_start + adj_start,
1011 adjust->vm_end);
1012 }
1013
1014 if (vma_iter_prealloc(vmi, vma))
1015 goto prealloc_fail;
1016
1017 init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
1018 VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
1019 vp.anon_vma != adjust->anon_vma);
1020
1021 vma_prepare(&vp);
1022 vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1023
1024 vma->vm_start = vma_start;
1025 vma->vm_end = vma_end;
1026 vma->vm_pgoff = vma_pgoff;
1027
1028 if (vma_expanded)
1029 vma_iter_store(vmi, vma);
1030
1031 if (adj_start) {
1032 adjust->vm_start += adj_start;
1033 adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1034 if (adj_start < 0) {
1035 WARN_ON(vma_expanded);
1036 vma_iter_store(vmi, next);
1037 }
1038 }
1039
1040 vma_complete(&vp, vmi, mm);
1041 khugepaged_enter_vma(res, vm_flags);
1042 return res;
1043
1044prealloc_fail:
1045 if (anon_dup)
1046 unlink_anon_vmas(anon_dup);
1047
1048anon_vma_fail:
1049 vma_iter_set(vmi, addr);
1050 vma_iter_load(vmi);
1051 return NULL;
1052}
1053
1054/*
1055 * Rough compatibility check to quickly see if it's even worth looking
1056 * at sharing an anon_vma.
1057 *
1058 * They need to have the same vm_file, and the flags can only differ
1059 * in things that mprotect may change.
1060 *
1061 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1062 * we can merge the two vma's. For example, we refuse to merge a vma if
1063 * there is a vm_ops->close() function, because that indicates that the
1064 * driver is doing some kind of reference counting. But that doesn't
1065 * really matter for the anon_vma sharing case.
1066 */
1067static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1068{
1069 return a->vm_end == b->vm_start &&
1070 mpol_equal(vma_policy(a), vma_policy(b)) &&
1071 a->vm_file == b->vm_file &&
1072 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1073 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1074}
1075
1076/*
1077 * Do some basic sanity checking to see if we can re-use the anon_vma
1078 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1079 * the same as 'old', the other will be the new one that is trying
1080 * to share the anon_vma.
1081 *
1082 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1083 * the anon_vma of 'old' is concurrently in the process of being set up
1084 * by another page fault trying to merge _that_. But that's ok: if it
1085 * is being set up, that automatically means that it will be a singleton
1086 * acceptable for merging, so we can do all of this optimistically. But
1087 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1088 *
1089 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1090 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1091 * is to return an anon_vma that is "complex" due to having gone through
1092 * a fork).
1093 *
1094 * We also make sure that the two vma's are compatible (adjacent,
1095 * and with the same memory policies). That's all stable, even with just
1096 * a read lock on the mmap_lock.
1097 */
1098static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1099{
1100 if (anon_vma_compatible(a, b)) {
1101 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1102
1103 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1104 return anon_vma;
1105 }
1106 return NULL;
1107}
1108
1109/*
1110 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1111 * neighbouring vmas for a suitable anon_vma, before it goes off
1112 * to allocate a new anon_vma. It checks because a repetitive
1113 * sequence of mprotects and faults may otherwise lead to distinct
1114 * anon_vmas being allocated, preventing vma merge in subsequent
1115 * mprotect.
1116 */
1117struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1118{
1119 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1120 struct anon_vma *anon_vma = NULL;
1121 struct vm_area_struct *prev, *next;
1122
1123 /* Try next first. */
1124 next = mas_walk(&mas);
1125 if (next) {
1126 anon_vma = reusable_anon_vma(next, vma, next);
1127 if (anon_vma)
1128 return anon_vma;
1129 }
1130
1131 prev = mas_prev(&mas, 0);
1132 VM_BUG_ON_VMA(prev != vma, vma);
1133 prev = mas_prev(&mas, 0);
1134 /* Try prev next. */
1135 if (prev)
1136 anon_vma = reusable_anon_vma(prev, prev, vma);
1137
1138 /*
1139 * We might reach here with anon_vma == NULL if we can't find
1140 * any reusable anon_vma.
1141 * There's no absolute need to look only at touching neighbours:
1142 * we could search further afield for "compatible" anon_vmas.
1143 * But it would probably just be a waste of time searching,
1144 * or lead to too many vmas hanging off the same anon_vma.
1145 * We're trying to allow mprotect remerging later on,
1146 * not trying to minimize memory used for anon_vmas.
1147 */
1148 return anon_vma;
1149}
1150
1151/*
1152 * If a hint addr is less than mmap_min_addr change hint to be as
1153 * low as possible but still greater than mmap_min_addr
1154 */
1155static inline unsigned long round_hint_to_min(unsigned long hint)
1156{
1157 hint &= PAGE_MASK;
1158 if (((void *)hint != NULL) &&
1159 (hint < mmap_min_addr))
1160 return PAGE_ALIGN(mmap_min_addr);
1161 return hint;
1162}
1163
1164bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
1165 unsigned long bytes)
1166{
1167 unsigned long locked_pages, limit_pages;
1168
1169 if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK))
1170 return true;
1171
1172 locked_pages = bytes >> PAGE_SHIFT;
1173 locked_pages += mm->locked_vm;
1174
1175 limit_pages = rlimit(RLIMIT_MEMLOCK);
1176 limit_pages >>= PAGE_SHIFT;
1177
1178 return locked_pages <= limit_pages;
1179}
1180
1181static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1182{
1183 if (S_ISREG(inode->i_mode))
1184 return MAX_LFS_FILESIZE;
1185
1186 if (S_ISBLK(inode->i_mode))
1187 return MAX_LFS_FILESIZE;
1188
1189 if (S_ISSOCK(inode->i_mode))
1190 return MAX_LFS_FILESIZE;
1191
1192 /* Special "we do even unsigned file positions" case */
1193 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1194 return 0;
1195
1196 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1197 return ULONG_MAX;
1198}
1199
1200static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1201 unsigned long pgoff, unsigned long len)
1202{
1203 u64 maxsize = file_mmap_size_max(file, inode);
1204
1205 if (maxsize && len > maxsize)
1206 return false;
1207 maxsize -= len;
1208 if (pgoff > maxsize >> PAGE_SHIFT)
1209 return false;
1210 return true;
1211}
1212
1213/*
1214 * The caller must write-lock current->mm->mmap_lock.
1215 */
1216unsigned long do_mmap(struct file *file, unsigned long addr,
1217 unsigned long len, unsigned long prot,
1218 unsigned long flags, vm_flags_t vm_flags,
1219 unsigned long pgoff, unsigned long *populate,
1220 struct list_head *uf)
1221{
1222 struct mm_struct *mm = current->mm;
1223 int pkey = 0;
1224
1225 *populate = 0;
1226
1227 if (!len)
1228 return -EINVAL;
1229
1230 /*
1231 * Does the application expect PROT_READ to imply PROT_EXEC?
1232 *
1233 * (the exception is when the underlying filesystem is noexec
1234 * mounted, in which case we don't add PROT_EXEC.)
1235 */
1236 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1237 if (!(file && path_noexec(&file->f_path)))
1238 prot |= PROT_EXEC;
1239
1240 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1241 if (flags & MAP_FIXED_NOREPLACE)
1242 flags |= MAP_FIXED;
1243
1244 if (!(flags & MAP_FIXED))
1245 addr = round_hint_to_min(addr);
1246
1247 /* Careful about overflows.. */
1248 len = PAGE_ALIGN(len);
1249 if (!len)
1250 return -ENOMEM;
1251
1252 /* offset overflow? */
1253 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1254 return -EOVERFLOW;
1255
1256 /* Too many mappings? */
1257 if (mm->map_count > sysctl_max_map_count)
1258 return -ENOMEM;
1259
1260 /* Obtain the address to map to. we verify (or select) it and ensure
1261 * that it represents a valid section of the address space.
1262 */
1263 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1264 if (IS_ERR_VALUE(addr))
1265 return addr;
1266
1267 if (flags & MAP_FIXED_NOREPLACE) {
1268 if (find_vma_intersection(mm, addr, addr + len))
1269 return -EEXIST;
1270 }
1271
1272 if (prot == PROT_EXEC) {
1273 pkey = execute_only_pkey(mm);
1274 if (pkey < 0)
1275 pkey = 0;
1276 }
1277
1278 /* Do simple checking here so the lower-level routines won't have
1279 * to. we assume access permissions have been handled by the open
1280 * of the memory object, so we don't do any here.
1281 */
1282 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1283 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1284
1285 if (flags & MAP_LOCKED)
1286 if (!can_do_mlock())
1287 return -EPERM;
1288
1289 if (!mlock_future_ok(mm, vm_flags, len))
1290 return -EAGAIN;
1291
1292 if (file) {
1293 struct inode *inode = file_inode(file);
1294 unsigned long flags_mask;
1295
1296 if (!file_mmap_ok(file, inode, pgoff, len))
1297 return -EOVERFLOW;
1298
1299 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1300
1301 switch (flags & MAP_TYPE) {
1302 case MAP_SHARED:
1303 /*
1304 * Force use of MAP_SHARED_VALIDATE with non-legacy
1305 * flags. E.g. MAP_SYNC is dangerous to use with
1306 * MAP_SHARED as you don't know which consistency model
1307 * you will get. We silently ignore unsupported flags
1308 * with MAP_SHARED to preserve backward compatibility.
1309 */
1310 flags &= LEGACY_MAP_MASK;
1311 fallthrough;
1312 case MAP_SHARED_VALIDATE:
1313 if (flags & ~flags_mask)
1314 return -EOPNOTSUPP;
1315 if (prot & PROT_WRITE) {
1316 if (!(file->f_mode & FMODE_WRITE))
1317 return -EACCES;
1318 if (IS_SWAPFILE(file->f_mapping->host))
1319 return -ETXTBSY;
1320 }
1321
1322 /*
1323 * Make sure we don't allow writing to an append-only
1324 * file..
1325 */
1326 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1327 return -EACCES;
1328
1329 vm_flags |= VM_SHARED | VM_MAYSHARE;
1330 if (!(file->f_mode & FMODE_WRITE))
1331 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1332 fallthrough;
1333 case MAP_PRIVATE:
1334 if (!(file->f_mode & FMODE_READ))
1335 return -EACCES;
1336 if (path_noexec(&file->f_path)) {
1337 if (vm_flags & VM_EXEC)
1338 return -EPERM;
1339 vm_flags &= ~VM_MAYEXEC;
1340 }
1341
1342 if (!file->f_op->mmap)
1343 return -ENODEV;
1344 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1345 return -EINVAL;
1346 break;
1347
1348 default:
1349 return -EINVAL;
1350 }
1351 } else {
1352 switch (flags & MAP_TYPE) {
1353 case MAP_SHARED:
1354 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1355 return -EINVAL;
1356 /*
1357 * Ignore pgoff.
1358 */
1359 pgoff = 0;
1360 vm_flags |= VM_SHARED | VM_MAYSHARE;
1361 break;
1362 case MAP_PRIVATE:
1363 /*
1364 * Set pgoff according to addr for anon_vma.
1365 */
1366 pgoff = addr >> PAGE_SHIFT;
1367 break;
1368 default:
1369 return -EINVAL;
1370 }
1371 }
1372
1373 /*
1374 * Set 'VM_NORESERVE' if we should not account for the
1375 * memory use of this mapping.
1376 */
1377 if (flags & MAP_NORESERVE) {
1378 /* We honor MAP_NORESERVE if allowed to overcommit */
1379 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1380 vm_flags |= VM_NORESERVE;
1381
1382 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1383 if (file && is_file_hugepages(file))
1384 vm_flags |= VM_NORESERVE;
1385 }
1386
1387 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1388 if (!IS_ERR_VALUE(addr) &&
1389 ((vm_flags & VM_LOCKED) ||
1390 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1391 *populate = len;
1392 return addr;
1393}
1394
1395unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1396 unsigned long prot, unsigned long flags,
1397 unsigned long fd, unsigned long pgoff)
1398{
1399 struct file *file = NULL;
1400 unsigned long retval;
1401
1402 if (!(flags & MAP_ANONYMOUS)) {
1403 audit_mmap_fd(fd, flags);
1404 file = fget(fd);
1405 if (!file)
1406 return -EBADF;
1407 if (is_file_hugepages(file)) {
1408 len = ALIGN(len, huge_page_size(hstate_file(file)));
1409 } else if (unlikely(flags & MAP_HUGETLB)) {
1410 retval = -EINVAL;
1411 goto out_fput;
1412 }
1413 } else if (flags & MAP_HUGETLB) {
1414 struct hstate *hs;
1415
1416 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1417 if (!hs)
1418 return -EINVAL;
1419
1420 len = ALIGN(len, huge_page_size(hs));
1421 /*
1422 * VM_NORESERVE is used because the reservations will be
1423 * taken when vm_ops->mmap() is called
1424 */
1425 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1426 VM_NORESERVE,
1427 HUGETLB_ANONHUGE_INODE,
1428 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1429 if (IS_ERR(file))
1430 return PTR_ERR(file);
1431 }
1432
1433 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1434out_fput:
1435 if (file)
1436 fput(file);
1437 return retval;
1438}
1439
1440SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1441 unsigned long, prot, unsigned long, flags,
1442 unsigned long, fd, unsigned long, pgoff)
1443{
1444 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1445}
1446
1447#ifdef __ARCH_WANT_SYS_OLD_MMAP
1448struct mmap_arg_struct {
1449 unsigned long addr;
1450 unsigned long len;
1451 unsigned long prot;
1452 unsigned long flags;
1453 unsigned long fd;
1454 unsigned long offset;
1455};
1456
1457SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1458{
1459 struct mmap_arg_struct a;
1460
1461 if (copy_from_user(&a, arg, sizeof(a)))
1462 return -EFAULT;
1463 if (offset_in_page(a.offset))
1464 return -EINVAL;
1465
1466 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1467 a.offset >> PAGE_SHIFT);
1468}
1469#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1470
1471static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops)
1472{
1473 return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite);
1474}
1475
1476static bool vma_is_shared_writable(struct vm_area_struct *vma)
1477{
1478 return (vma->vm_flags & (VM_WRITE | VM_SHARED)) ==
1479 (VM_WRITE | VM_SHARED);
1480}
1481
1482static bool vma_fs_can_writeback(struct vm_area_struct *vma)
1483{
1484 /* No managed pages to writeback. */
1485 if (vma->vm_flags & VM_PFNMAP)
1486 return false;
1487
1488 return vma->vm_file && vma->vm_file->f_mapping &&
1489 mapping_can_writeback(vma->vm_file->f_mapping);
1490}
1491
1492/*
1493 * Does this VMA require the underlying folios to have their dirty state
1494 * tracked?
1495 */
1496bool vma_needs_dirty_tracking(struct vm_area_struct *vma)
1497{
1498 /* Only shared, writable VMAs require dirty tracking. */
1499 if (!vma_is_shared_writable(vma))
1500 return false;
1501
1502 /* Does the filesystem need to be notified? */
1503 if (vm_ops_needs_writenotify(vma->vm_ops))
1504 return true;
1505
1506 /*
1507 * Even if the filesystem doesn't indicate a need for writenotify, if it
1508 * can writeback, dirty tracking is still required.
1509 */
1510 return vma_fs_can_writeback(vma);
1511}
1512
1513/*
1514 * Some shared mappings will want the pages marked read-only
1515 * to track write events. If so, we'll downgrade vm_page_prot
1516 * to the private version (using protection_map[] without the
1517 * VM_SHARED bit).
1518 */
1519int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1520{
1521 /* If it was private or non-writable, the write bit is already clear */
1522 if (!vma_is_shared_writable(vma))
1523 return 0;
1524
1525 /* The backer wishes to know when pages are first written to? */
1526 if (vm_ops_needs_writenotify(vma->vm_ops))
1527 return 1;
1528
1529 /* The open routine did something to the protections that pgprot_modify
1530 * won't preserve? */
1531 if (pgprot_val(vm_page_prot) !=
1532 pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags)))
1533 return 0;
1534
1535 /*
1536 * Do we need to track softdirty? hugetlb does not support softdirty
1537 * tracking yet.
1538 */
1539 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1540 return 1;
1541
1542 /* Do we need write faults for uffd-wp tracking? */
1543 if (userfaultfd_wp(vma))
1544 return 1;
1545
1546 /* Can the mapping track the dirty pages? */
1547 return vma_fs_can_writeback(vma);
1548}
1549
1550/*
1551 * We account for memory if it's a private writeable mapping,
1552 * not hugepages and VM_NORESERVE wasn't set.
1553 */
1554static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1555{
1556 /*
1557 * hugetlb has its own accounting separate from the core VM
1558 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1559 */
1560 if (file && is_file_hugepages(file))
1561 return 0;
1562
1563 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1564}
1565
1566/**
1567 * unmapped_area() - Find an area between the low_limit and the high_limit with
1568 * the correct alignment and offset, all from @info. Note: current->mm is used
1569 * for the search.
1570 *
1571 * @info: The unmapped area information including the range [low_limit -
1572 * high_limit), the alignment offset and mask.
1573 *
1574 * Return: A memory address or -ENOMEM.
1575 */
1576static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1577{
1578 unsigned long length, gap;
1579 unsigned long low_limit, high_limit;
1580 struct vm_area_struct *tmp;
1581
1582 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1583
1584 /* Adjust search length to account for worst case alignment overhead */
1585 length = info->length + info->align_mask;
1586 if (length < info->length)
1587 return -ENOMEM;
1588
1589 low_limit = info->low_limit;
1590 if (low_limit < mmap_min_addr)
1591 low_limit = mmap_min_addr;
1592 high_limit = info->high_limit;
1593retry:
1594 if (mas_empty_area(&mas, low_limit, high_limit - 1, length))
1595 return -ENOMEM;
1596
1597 gap = mas.index;
1598 gap += (info->align_offset - gap) & info->align_mask;
1599 tmp = mas_next(&mas, ULONG_MAX);
1600 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1601 if (vm_start_gap(tmp) < gap + length - 1) {
1602 low_limit = tmp->vm_end;
1603 mas_reset(&mas);
1604 goto retry;
1605 }
1606 } else {
1607 tmp = mas_prev(&mas, 0);
1608 if (tmp && vm_end_gap(tmp) > gap) {
1609 low_limit = vm_end_gap(tmp);
1610 mas_reset(&mas);
1611 goto retry;
1612 }
1613 }
1614
1615 return gap;
1616}
1617
1618/**
1619 * unmapped_area_topdown() - Find an area between the low_limit and the
1620 * high_limit with the correct alignment and offset at the highest available
1621 * address, all from @info. Note: current->mm is used for the search.
1622 *
1623 * @info: The unmapped area information including the range [low_limit -
1624 * high_limit), the alignment offset and mask.
1625 *
1626 * Return: A memory address or -ENOMEM.
1627 */
1628static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1629{
1630 unsigned long length, gap, gap_end;
1631 unsigned long low_limit, high_limit;
1632 struct vm_area_struct *tmp;
1633
1634 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1635 /* Adjust search length to account for worst case alignment overhead */
1636 length = info->length + info->align_mask;
1637 if (length < info->length)
1638 return -ENOMEM;
1639
1640 low_limit = info->low_limit;
1641 if (low_limit < mmap_min_addr)
1642 low_limit = mmap_min_addr;
1643 high_limit = info->high_limit;
1644retry:
1645 if (mas_empty_area_rev(&mas, low_limit, high_limit - 1, length))
1646 return -ENOMEM;
1647
1648 gap = mas.last + 1 - info->length;
1649 gap -= (gap - info->align_offset) & info->align_mask;
1650 gap_end = mas.last;
1651 tmp = mas_next(&mas, ULONG_MAX);
1652 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1653 if (vm_start_gap(tmp) <= gap_end) {
1654 high_limit = vm_start_gap(tmp);
1655 mas_reset(&mas);
1656 goto retry;
1657 }
1658 } else {
1659 tmp = mas_prev(&mas, 0);
1660 if (tmp && vm_end_gap(tmp) > gap) {
1661 high_limit = tmp->vm_start;
1662 mas_reset(&mas);
1663 goto retry;
1664 }
1665 }
1666
1667 return gap;
1668}
1669
1670/*
1671 * Search for an unmapped address range.
1672 *
1673 * We are looking for a range that:
1674 * - does not intersect with any VMA;
1675 * - is contained within the [low_limit, high_limit) interval;
1676 * - is at least the desired size.
1677 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1678 */
1679unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1680{
1681 unsigned long addr;
1682
1683 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1684 addr = unmapped_area_topdown(info);
1685 else
1686 addr = unmapped_area(info);
1687
1688 trace_vm_unmapped_area(addr, info);
1689 return addr;
1690}
1691
1692/* Get an address range which is currently unmapped.
1693 * For shmat() with addr=0.
1694 *
1695 * Ugly calling convention alert:
1696 * Return value with the low bits set means error value,
1697 * ie
1698 * if (ret & ~PAGE_MASK)
1699 * error = ret;
1700 *
1701 * This function "knows" that -ENOMEM has the bits set.
1702 */
1703unsigned long
1704generic_get_unmapped_area(struct file *filp, unsigned long addr,
1705 unsigned long len, unsigned long pgoff,
1706 unsigned long flags)
1707{
1708 struct mm_struct *mm = current->mm;
1709 struct vm_area_struct *vma, *prev;
1710 struct vm_unmapped_area_info info;
1711 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1712
1713 if (len > mmap_end - mmap_min_addr)
1714 return -ENOMEM;
1715
1716 if (flags & MAP_FIXED)
1717 return addr;
1718
1719 if (addr) {
1720 addr = PAGE_ALIGN(addr);
1721 vma = find_vma_prev(mm, addr, &prev);
1722 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1723 (!vma || addr + len <= vm_start_gap(vma)) &&
1724 (!prev || addr >= vm_end_gap(prev)))
1725 return addr;
1726 }
1727
1728 info.flags = 0;
1729 info.length = len;
1730 info.low_limit = mm->mmap_base;
1731 info.high_limit = mmap_end;
1732 info.align_mask = 0;
1733 info.align_offset = 0;
1734 return vm_unmapped_area(&info);
1735}
1736
1737#ifndef HAVE_ARCH_UNMAPPED_AREA
1738unsigned long
1739arch_get_unmapped_area(struct file *filp, unsigned long addr,
1740 unsigned long len, unsigned long pgoff,
1741 unsigned long flags)
1742{
1743 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1744}
1745#endif
1746
1747/*
1748 * This mmap-allocator allocates new areas top-down from below the
1749 * stack's low limit (the base):
1750 */
1751unsigned long
1752generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1753 unsigned long len, unsigned long pgoff,
1754 unsigned long flags)
1755{
1756 struct vm_area_struct *vma, *prev;
1757 struct mm_struct *mm = current->mm;
1758 struct vm_unmapped_area_info info;
1759 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1760
1761 /* requested length too big for entire address space */
1762 if (len > mmap_end - mmap_min_addr)
1763 return -ENOMEM;
1764
1765 if (flags & MAP_FIXED)
1766 return addr;
1767
1768 /* requesting a specific address */
1769 if (addr) {
1770 addr = PAGE_ALIGN(addr);
1771 vma = find_vma_prev(mm, addr, &prev);
1772 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1773 (!vma || addr + len <= vm_start_gap(vma)) &&
1774 (!prev || addr >= vm_end_gap(prev)))
1775 return addr;
1776 }
1777
1778 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1779 info.length = len;
1780 info.low_limit = PAGE_SIZE;
1781 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1782 info.align_mask = 0;
1783 info.align_offset = 0;
1784 addr = vm_unmapped_area(&info);
1785
1786 /*
1787 * A failed mmap() very likely causes application failure,
1788 * so fall back to the bottom-up function here. This scenario
1789 * can happen with large stack limits and large mmap()
1790 * allocations.
1791 */
1792 if (offset_in_page(addr)) {
1793 VM_BUG_ON(addr != -ENOMEM);
1794 info.flags = 0;
1795 info.low_limit = TASK_UNMAPPED_BASE;
1796 info.high_limit = mmap_end;
1797 addr = vm_unmapped_area(&info);
1798 }
1799
1800 return addr;
1801}
1802
1803#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1804unsigned long
1805arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1806 unsigned long len, unsigned long pgoff,
1807 unsigned long flags)
1808{
1809 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1810}
1811#endif
1812
1813unsigned long
1814get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1815 unsigned long pgoff, unsigned long flags)
1816{
1817 unsigned long (*get_area)(struct file *, unsigned long,
1818 unsigned long, unsigned long, unsigned long);
1819
1820 unsigned long error = arch_mmap_check(addr, len, flags);
1821 if (error)
1822 return error;
1823
1824 /* Careful about overflows.. */
1825 if (len > TASK_SIZE)
1826 return -ENOMEM;
1827
1828 get_area = current->mm->get_unmapped_area;
1829 if (file) {
1830 if (file->f_op->get_unmapped_area)
1831 get_area = file->f_op->get_unmapped_area;
1832 } else if (flags & MAP_SHARED) {
1833 /*
1834 * mmap_region() will call shmem_zero_setup() to create a file,
1835 * so use shmem's get_unmapped_area in case it can be huge.
1836 */
1837 get_area = shmem_get_unmapped_area;
1838 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1839 /* Ensures that larger anonymous mappings are THP aligned. */
1840 get_area = thp_get_unmapped_area;
1841 }
1842
1843 /* Always treat pgoff as zero for anonymous memory. */
1844 if (!file)
1845 pgoff = 0;
1846
1847 addr = get_area(file, addr, len, pgoff, flags);
1848 if (IS_ERR_VALUE(addr))
1849 return addr;
1850
1851 if (addr > TASK_SIZE - len)
1852 return -ENOMEM;
1853 if (offset_in_page(addr))
1854 return -EINVAL;
1855
1856 error = security_mmap_addr(addr);
1857 return error ? error : addr;
1858}
1859
1860EXPORT_SYMBOL(get_unmapped_area);
1861
1862/**
1863 * find_vma_intersection() - Look up the first VMA which intersects the interval
1864 * @mm: The process address space.
1865 * @start_addr: The inclusive start user address.
1866 * @end_addr: The exclusive end user address.
1867 *
1868 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1869 * start_addr < end_addr.
1870 */
1871struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1872 unsigned long start_addr,
1873 unsigned long end_addr)
1874{
1875 unsigned long index = start_addr;
1876
1877 mmap_assert_locked(mm);
1878 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1879}
1880EXPORT_SYMBOL(find_vma_intersection);
1881
1882/**
1883 * find_vma() - Find the VMA for a given address, or the next VMA.
1884 * @mm: The mm_struct to check
1885 * @addr: The address
1886 *
1887 * Returns: The VMA associated with addr, or the next VMA.
1888 * May return %NULL in the case of no VMA at addr or above.
1889 */
1890struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1891{
1892 unsigned long index = addr;
1893
1894 mmap_assert_locked(mm);
1895 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1896}
1897EXPORT_SYMBOL(find_vma);
1898
1899/**
1900 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1901 * set %pprev to the previous VMA, if any.
1902 * @mm: The mm_struct to check
1903 * @addr: The address
1904 * @pprev: The pointer to set to the previous VMA
1905 *
1906 * Note that RCU lock is missing here since the external mmap_lock() is used
1907 * instead.
1908 *
1909 * Returns: The VMA associated with @addr, or the next vma.
1910 * May return %NULL in the case of no vma at addr or above.
1911 */
1912struct vm_area_struct *
1913find_vma_prev(struct mm_struct *mm, unsigned long addr,
1914 struct vm_area_struct **pprev)
1915{
1916 struct vm_area_struct *vma;
1917 MA_STATE(mas, &mm->mm_mt, addr, addr);
1918
1919 vma = mas_walk(&mas);
1920 *pprev = mas_prev(&mas, 0);
1921 if (!vma)
1922 vma = mas_next(&mas, ULONG_MAX);
1923 return vma;
1924}
1925
1926/*
1927 * Verify that the stack growth is acceptable and
1928 * update accounting. This is shared with both the
1929 * grow-up and grow-down cases.
1930 */
1931static int acct_stack_growth(struct vm_area_struct *vma,
1932 unsigned long size, unsigned long grow)
1933{
1934 struct mm_struct *mm = vma->vm_mm;
1935 unsigned long new_start;
1936
1937 /* address space limit tests */
1938 if (!may_expand_vm(mm, vma->vm_flags, grow))
1939 return -ENOMEM;
1940
1941 /* Stack limit test */
1942 if (size > rlimit(RLIMIT_STACK))
1943 return -ENOMEM;
1944
1945 /* mlock limit tests */
1946 if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT))
1947 return -ENOMEM;
1948
1949 /* Check to ensure the stack will not grow into a hugetlb-only region */
1950 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1951 vma->vm_end - size;
1952 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1953 return -EFAULT;
1954
1955 /*
1956 * Overcommit.. This must be the final test, as it will
1957 * update security statistics.
1958 */
1959 if (security_vm_enough_memory_mm(mm, grow))
1960 return -ENOMEM;
1961
1962 return 0;
1963}
1964
1965#if defined(CONFIG_STACK_GROWSUP)
1966/*
1967 * PA-RISC uses this for its stack.
1968 * vma is the last one with address > vma->vm_end. Have to extend vma.
1969 */
1970static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1971{
1972 struct mm_struct *mm = vma->vm_mm;
1973 struct vm_area_struct *next;
1974 unsigned long gap_addr;
1975 int error = 0;
1976 MA_STATE(mas, &mm->mm_mt, vma->vm_start, address);
1977
1978 if (!(vma->vm_flags & VM_GROWSUP))
1979 return -EFAULT;
1980
1981 /* Guard against exceeding limits of the address space. */
1982 address &= PAGE_MASK;
1983 if (address >= (TASK_SIZE & PAGE_MASK))
1984 return -ENOMEM;
1985 address += PAGE_SIZE;
1986
1987 /* Enforce stack_guard_gap */
1988 gap_addr = address + stack_guard_gap;
1989
1990 /* Guard against overflow */
1991 if (gap_addr < address || gap_addr > TASK_SIZE)
1992 gap_addr = TASK_SIZE;
1993
1994 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1995 if (next && vma_is_accessible(next)) {
1996 if (!(next->vm_flags & VM_GROWSUP))
1997 return -ENOMEM;
1998 /* Check that both stack segments have the same anon_vma? */
1999 }
2000
2001 if (next)
2002 mas_prev_range(&mas, address);
2003
2004 __mas_set_range(&mas, vma->vm_start, address - 1);
2005 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2006 return -ENOMEM;
2007
2008 /* We must make sure the anon_vma is allocated. */
2009 if (unlikely(anon_vma_prepare(vma))) {
2010 mas_destroy(&mas);
2011 return -ENOMEM;
2012 }
2013
2014 /* Lock the VMA before expanding to prevent concurrent page faults */
2015 vma_start_write(vma);
2016 /*
2017 * vma->vm_start/vm_end cannot change under us because the caller
2018 * is required to hold the mmap_lock in read mode. We need the
2019 * anon_vma lock to serialize against concurrent expand_stacks.
2020 */
2021 anon_vma_lock_write(vma->anon_vma);
2022
2023 /* Somebody else might have raced and expanded it already */
2024 if (address > vma->vm_end) {
2025 unsigned long size, grow;
2026
2027 size = address - vma->vm_start;
2028 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2029
2030 error = -ENOMEM;
2031 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2032 error = acct_stack_growth(vma, size, grow);
2033 if (!error) {
2034 /*
2035 * We only hold a shared mmap_lock lock here, so
2036 * we need to protect against concurrent vma
2037 * expansions. anon_vma_lock_write() doesn't
2038 * help here, as we don't guarantee that all
2039 * growable vmas in a mm share the same root
2040 * anon vma. So, we reuse mm->page_table_lock
2041 * to guard against concurrent vma expansions.
2042 */
2043 spin_lock(&mm->page_table_lock);
2044 if (vma->vm_flags & VM_LOCKED)
2045 mm->locked_vm += grow;
2046 vm_stat_account(mm, vma->vm_flags, grow);
2047 anon_vma_interval_tree_pre_update_vma(vma);
2048 vma->vm_end = address;
2049 /* Overwrite old entry in mtree. */
2050 mas_store_prealloc(&mas, vma);
2051 anon_vma_interval_tree_post_update_vma(vma);
2052 spin_unlock(&mm->page_table_lock);
2053
2054 perf_event_mmap(vma);
2055 }
2056 }
2057 }
2058 anon_vma_unlock_write(vma->anon_vma);
2059 khugepaged_enter_vma(vma, vma->vm_flags);
2060 mas_destroy(&mas);
2061 validate_mm(mm);
2062 return error;
2063}
2064#endif /* CONFIG_STACK_GROWSUP */
2065
2066/*
2067 * vma is the first one with address < vma->vm_start. Have to extend vma.
2068 * mmap_lock held for writing.
2069 */
2070int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2071{
2072 struct mm_struct *mm = vma->vm_mm;
2073 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2074 struct vm_area_struct *prev;
2075 int error = 0;
2076
2077 if (!(vma->vm_flags & VM_GROWSDOWN))
2078 return -EFAULT;
2079
2080 address &= PAGE_MASK;
2081 if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
2082 return -EPERM;
2083
2084 /* Enforce stack_guard_gap */
2085 prev = mas_prev(&mas, 0);
2086 /* Check that both stack segments have the same anon_vma? */
2087 if (prev) {
2088 if (!(prev->vm_flags & VM_GROWSDOWN) &&
2089 vma_is_accessible(prev) &&
2090 (address - prev->vm_end < stack_guard_gap))
2091 return -ENOMEM;
2092 }
2093
2094 if (prev)
2095 mas_next_range(&mas, vma->vm_start);
2096
2097 __mas_set_range(&mas, address, vma->vm_end - 1);
2098 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2099 return -ENOMEM;
2100
2101 /* We must make sure the anon_vma is allocated. */
2102 if (unlikely(anon_vma_prepare(vma))) {
2103 mas_destroy(&mas);
2104 return -ENOMEM;
2105 }
2106
2107 /* Lock the VMA before expanding to prevent concurrent page faults */
2108 vma_start_write(vma);
2109 /*
2110 * vma->vm_start/vm_end cannot change under us because the caller
2111 * is required to hold the mmap_lock in read mode. We need the
2112 * anon_vma lock to serialize against concurrent expand_stacks.
2113 */
2114 anon_vma_lock_write(vma->anon_vma);
2115
2116 /* Somebody else might have raced and expanded it already */
2117 if (address < vma->vm_start) {
2118 unsigned long size, grow;
2119
2120 size = vma->vm_end - address;
2121 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2122
2123 error = -ENOMEM;
2124 if (grow <= vma->vm_pgoff) {
2125 error = acct_stack_growth(vma, size, grow);
2126 if (!error) {
2127 /*
2128 * We only hold a shared mmap_lock lock here, so
2129 * we need to protect against concurrent vma
2130 * expansions. anon_vma_lock_write() doesn't
2131 * help here, as we don't guarantee that all
2132 * growable vmas in a mm share the same root
2133 * anon vma. So, we reuse mm->page_table_lock
2134 * to guard against concurrent vma expansions.
2135 */
2136 spin_lock(&mm->page_table_lock);
2137 if (vma->vm_flags & VM_LOCKED)
2138 mm->locked_vm += grow;
2139 vm_stat_account(mm, vma->vm_flags, grow);
2140 anon_vma_interval_tree_pre_update_vma(vma);
2141 vma->vm_start = address;
2142 vma->vm_pgoff -= grow;
2143 /* Overwrite old entry in mtree. */
2144 mas_store_prealloc(&mas, vma);
2145 anon_vma_interval_tree_post_update_vma(vma);
2146 spin_unlock(&mm->page_table_lock);
2147
2148 perf_event_mmap(vma);
2149 }
2150 }
2151 }
2152 anon_vma_unlock_write(vma->anon_vma);
2153 khugepaged_enter_vma(vma, vma->vm_flags);
2154 mas_destroy(&mas);
2155 validate_mm(mm);
2156 return error;
2157}
2158
2159/* enforced gap between the expanding stack and other mappings. */
2160unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2161
2162static int __init cmdline_parse_stack_guard_gap(char *p)
2163{
2164 unsigned long val;
2165 char *endptr;
2166
2167 val = simple_strtoul(p, &endptr, 10);
2168 if (!*endptr)
2169 stack_guard_gap = val << PAGE_SHIFT;
2170
2171 return 1;
2172}
2173__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2174
2175#ifdef CONFIG_STACK_GROWSUP
2176int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2177{
2178 return expand_upwards(vma, address);
2179}
2180
2181struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2182{
2183 struct vm_area_struct *vma, *prev;
2184
2185 addr &= PAGE_MASK;
2186 vma = find_vma_prev(mm, addr, &prev);
2187 if (vma && (vma->vm_start <= addr))
2188 return vma;
2189 if (!prev)
2190 return NULL;
2191 if (expand_stack_locked(prev, addr))
2192 return NULL;
2193 if (prev->vm_flags & VM_LOCKED)
2194 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2195 return prev;
2196}
2197#else
2198int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2199{
2200 return expand_downwards(vma, address);
2201}
2202
2203struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2204{
2205 struct vm_area_struct *vma;
2206 unsigned long start;
2207
2208 addr &= PAGE_MASK;
2209 vma = find_vma(mm, addr);
2210 if (!vma)
2211 return NULL;
2212 if (vma->vm_start <= addr)
2213 return vma;
2214 start = vma->vm_start;
2215 if (expand_stack_locked(vma, addr))
2216 return NULL;
2217 if (vma->vm_flags & VM_LOCKED)
2218 populate_vma_page_range(vma, addr, start, NULL);
2219 return vma;
2220}
2221#endif
2222
2223#if defined(CONFIG_STACK_GROWSUP)
2224
2225#define vma_expand_up(vma,addr) expand_upwards(vma, addr)
2226#define vma_expand_down(vma, addr) (-EFAULT)
2227
2228#else
2229
2230#define vma_expand_up(vma,addr) (-EFAULT)
2231#define vma_expand_down(vma, addr) expand_downwards(vma, addr)
2232
2233#endif
2234
2235/*
2236 * expand_stack(): legacy interface for page faulting. Don't use unless
2237 * you have to.
2238 *
2239 * This is called with the mm locked for reading, drops the lock, takes
2240 * the lock for writing, tries to look up a vma again, expands it if
2241 * necessary, and downgrades the lock to reading again.
2242 *
2243 * If no vma is found or it can't be expanded, it returns NULL and has
2244 * dropped the lock.
2245 */
2246struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
2247{
2248 struct vm_area_struct *vma, *prev;
2249
2250 mmap_read_unlock(mm);
2251 if (mmap_write_lock_killable(mm))
2252 return NULL;
2253
2254 vma = find_vma_prev(mm, addr, &prev);
2255 if (vma && vma->vm_start <= addr)
2256 goto success;
2257
2258 if (prev && !vma_expand_up(prev, addr)) {
2259 vma = prev;
2260 goto success;
2261 }
2262
2263 if (vma && !vma_expand_down(vma, addr))
2264 goto success;
2265
2266 mmap_write_unlock(mm);
2267 return NULL;
2268
2269success:
2270 mmap_write_downgrade(mm);
2271 return vma;
2272}
2273
2274/*
2275 * Ok - we have the memory areas we should free on a maple tree so release them,
2276 * and do the vma updates.
2277 *
2278 * Called with the mm semaphore held.
2279 */
2280static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2281{
2282 unsigned long nr_accounted = 0;
2283 struct vm_area_struct *vma;
2284
2285 /* Update high watermark before we lower total_vm */
2286 update_hiwater_vm(mm);
2287 mas_for_each(mas, vma, ULONG_MAX) {
2288 long nrpages = vma_pages(vma);
2289
2290 if (vma->vm_flags & VM_ACCOUNT)
2291 nr_accounted += nrpages;
2292 vm_stat_account(mm, vma->vm_flags, -nrpages);
2293 remove_vma(vma, false);
2294 }
2295 vm_unacct_memory(nr_accounted);
2296}
2297
2298/*
2299 * Get rid of page table information in the indicated region.
2300 *
2301 * Called with the mm semaphore held.
2302 */
2303static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
2304 struct vm_area_struct *vma, struct vm_area_struct *prev,
2305 struct vm_area_struct *next, unsigned long start,
2306 unsigned long end, unsigned long tree_end, bool mm_wr_locked)
2307{
2308 struct mmu_gather tlb;
2309 unsigned long mt_start = mas->index;
2310
2311 lru_add_drain();
2312 tlb_gather_mmu(&tlb, mm);
2313 update_hiwater_rss(mm);
2314 unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
2315 mas_set(mas, mt_start);
2316 free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2317 next ? next->vm_start : USER_PGTABLES_CEILING,
2318 mm_wr_locked);
2319 tlb_finish_mmu(&tlb);
2320}
2321
2322/*
2323 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2324 * has already been checked or doesn't make sense to fail.
2325 * VMA Iterator will point to the end VMA.
2326 */
2327static int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2328 unsigned long addr, int new_below)
2329{
2330 struct vma_prepare vp;
2331 struct vm_area_struct *new;
2332 int err;
2333
2334 WARN_ON(vma->vm_start >= addr);
2335 WARN_ON(vma->vm_end <= addr);
2336
2337 if (vma->vm_ops && vma->vm_ops->may_split) {
2338 err = vma->vm_ops->may_split(vma, addr);
2339 if (err)
2340 return err;
2341 }
2342
2343 new = vm_area_dup(vma);
2344 if (!new)
2345 return -ENOMEM;
2346
2347 if (new_below) {
2348 new->vm_end = addr;
2349 } else {
2350 new->vm_start = addr;
2351 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2352 }
2353
2354 err = -ENOMEM;
2355 vma_iter_config(vmi, new->vm_start, new->vm_end);
2356 if (vma_iter_prealloc(vmi, new))
2357 goto out_free_vma;
2358
2359 err = vma_dup_policy(vma, new);
2360 if (err)
2361 goto out_free_vmi;
2362
2363 err = anon_vma_clone(new, vma);
2364 if (err)
2365 goto out_free_mpol;
2366
2367 if (new->vm_file)
2368 get_file(new->vm_file);
2369
2370 if (new->vm_ops && new->vm_ops->open)
2371 new->vm_ops->open(new);
2372
2373 vma_start_write(vma);
2374 vma_start_write(new);
2375
2376 init_vma_prep(&vp, vma);
2377 vp.insert = new;
2378 vma_prepare(&vp);
2379 vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
2380
2381 if (new_below) {
2382 vma->vm_start = addr;
2383 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
2384 } else {
2385 vma->vm_end = addr;
2386 }
2387
2388 /* vma_complete stores the new vma */
2389 vma_complete(&vp, vmi, vma->vm_mm);
2390
2391 /* Success. */
2392 if (new_below)
2393 vma_next(vmi);
2394 return 0;
2395
2396out_free_mpol:
2397 mpol_put(vma_policy(new));
2398out_free_vmi:
2399 vma_iter_free(vmi);
2400out_free_vma:
2401 vm_area_free(new);
2402 return err;
2403}
2404
2405/*
2406 * Split a vma into two pieces at address 'addr', a new vma is allocated
2407 * either for the first part or the tail.
2408 */
2409static int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2410 unsigned long addr, int new_below)
2411{
2412 if (vma->vm_mm->map_count >= sysctl_max_map_count)
2413 return -ENOMEM;
2414
2415 return __split_vma(vmi, vma, addr, new_below);
2416}
2417
2418/*
2419 * We are about to modify one or multiple of a VMA's flags, policy, userfaultfd
2420 * context and anonymous VMA name within the range [start, end).
2421 *
2422 * As a result, we might be able to merge the newly modified VMA range with an
2423 * adjacent VMA with identical properties.
2424 *
2425 * If no merge is possible and the range does not span the entirety of the VMA,
2426 * we then need to split the VMA to accommodate the change.
2427 *
2428 * The function returns either the merged VMA, the original VMA if a split was
2429 * required instead, or an error if the split failed.
2430 */
2431struct vm_area_struct *vma_modify(struct vma_iterator *vmi,
2432 struct vm_area_struct *prev,
2433 struct vm_area_struct *vma,
2434 unsigned long start, unsigned long end,
2435 unsigned long vm_flags,
2436 struct mempolicy *policy,
2437 struct vm_userfaultfd_ctx uffd_ctx,
2438 struct anon_vma_name *anon_name)
2439{
2440 pgoff_t pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
2441 struct vm_area_struct *merged;
2442
2443 merged = vma_merge(vmi, vma->vm_mm, prev, start, end, vm_flags,
2444 vma->anon_vma, vma->vm_file, pgoff, policy,
2445 uffd_ctx, anon_name);
2446 if (merged)
2447 return merged;
2448
2449 if (vma->vm_start < start) {
2450 int err = split_vma(vmi, vma, start, 1);
2451
2452 if (err)
2453 return ERR_PTR(err);
2454 }
2455
2456 if (vma->vm_end > end) {
2457 int err = split_vma(vmi, vma, end, 0);
2458
2459 if (err)
2460 return ERR_PTR(err);
2461 }
2462
2463 return vma;
2464}
2465
2466/*
2467 * Attempt to merge a newly mapped VMA with those adjacent to it. The caller
2468 * must ensure that [start, end) does not overlap any existing VMA.
2469 */
2470static struct vm_area_struct
2471*vma_merge_new_vma(struct vma_iterator *vmi, struct vm_area_struct *prev,
2472 struct vm_area_struct *vma, unsigned long start,
2473 unsigned long end, pgoff_t pgoff)
2474{
2475 return vma_merge(vmi, vma->vm_mm, prev, start, end, vma->vm_flags,
2476 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
2477 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2478}
2479
2480/*
2481 * Expand vma by delta bytes, potentially merging with an immediately adjacent
2482 * VMA with identical properties.
2483 */
2484struct vm_area_struct *vma_merge_extend(struct vma_iterator *vmi,
2485 struct vm_area_struct *vma,
2486 unsigned long delta)
2487{
2488 pgoff_t pgoff = vma->vm_pgoff + vma_pages(vma);
2489
2490 /* vma is specified as prev, so case 1 or 2 will apply. */
2491 return vma_merge(vmi, vma->vm_mm, vma, vma->vm_end, vma->vm_end + delta,
2492 vma->vm_flags, vma->anon_vma, vma->vm_file, pgoff,
2493 vma_policy(vma), vma->vm_userfaultfd_ctx,
2494 anon_vma_name(vma));
2495}
2496
2497/*
2498 * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
2499 * @vmi: The vma iterator
2500 * @vma: The starting vm_area_struct
2501 * @mm: The mm_struct
2502 * @start: The aligned start address to munmap.
2503 * @end: The aligned end address to munmap.
2504 * @uf: The userfaultfd list_head
2505 * @unlock: Set to true to drop the mmap_lock. unlocking only happens on
2506 * success.
2507 *
2508 * Return: 0 on success and drops the lock if so directed, error and leaves the
2509 * lock held otherwise.
2510 */
2511static int
2512do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2513 struct mm_struct *mm, unsigned long start,
2514 unsigned long end, struct list_head *uf, bool unlock)
2515{
2516 struct vm_area_struct *prev, *next = NULL;
2517 struct maple_tree mt_detach;
2518 int count = 0;
2519 int error = -ENOMEM;
2520 unsigned long locked_vm = 0;
2521 MA_STATE(mas_detach, &mt_detach, 0, 0);
2522 mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
2523 mt_on_stack(mt_detach);
2524
2525 /*
2526 * If we need to split any vma, do it now to save pain later.
2527 *
2528 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2529 * unmapped vm_area_struct will remain in use: so lower split_vma
2530 * places tmp vma above, and higher split_vma places tmp vma below.
2531 */
2532
2533 /* Does it split the first one? */
2534 if (start > vma->vm_start) {
2535
2536 /*
2537 * Make sure that map_count on return from munmap() will
2538 * not exceed its limit; but let map_count go just above
2539 * its limit temporarily, to help free resources as expected.
2540 */
2541 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2542 goto map_count_exceeded;
2543
2544 error = __split_vma(vmi, vma, start, 1);
2545 if (error)
2546 goto start_split_failed;
2547 }
2548
2549 /*
2550 * Detach a range of VMAs from the mm. Using next as a temp variable as
2551 * it is always overwritten.
2552 */
2553 next = vma;
2554 do {
2555 /* Does it split the end? */
2556 if (next->vm_end > end) {
2557 error = __split_vma(vmi, next, end, 0);
2558 if (error)
2559 goto end_split_failed;
2560 }
2561 vma_start_write(next);
2562 mas_set(&mas_detach, count);
2563 error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
2564 if (error)
2565 goto munmap_gather_failed;
2566 vma_mark_detached(next, true);
2567 if (next->vm_flags & VM_LOCKED)
2568 locked_vm += vma_pages(next);
2569
2570 count++;
2571 if (unlikely(uf)) {
2572 /*
2573 * If userfaultfd_unmap_prep returns an error the vmas
2574 * will remain split, but userland will get a
2575 * highly unexpected error anyway. This is no
2576 * different than the case where the first of the two
2577 * __split_vma fails, but we don't undo the first
2578 * split, despite we could. This is unlikely enough
2579 * failure that it's not worth optimizing it for.
2580 */
2581 error = userfaultfd_unmap_prep(next, start, end, uf);
2582
2583 if (error)
2584 goto userfaultfd_error;
2585 }
2586#ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2587 BUG_ON(next->vm_start < start);
2588 BUG_ON(next->vm_start > end);
2589#endif
2590 } for_each_vma_range(*vmi, next, end);
2591
2592#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2593 /* Make sure no VMAs are about to be lost. */
2594 {
2595 MA_STATE(test, &mt_detach, 0, 0);
2596 struct vm_area_struct *vma_mas, *vma_test;
2597 int test_count = 0;
2598
2599 vma_iter_set(vmi, start);
2600 rcu_read_lock();
2601 vma_test = mas_find(&test, count - 1);
2602 for_each_vma_range(*vmi, vma_mas, end) {
2603 BUG_ON(vma_mas != vma_test);
2604 test_count++;
2605 vma_test = mas_next(&test, count - 1);
2606 }
2607 rcu_read_unlock();
2608 BUG_ON(count != test_count);
2609 }
2610#endif
2611
2612 while (vma_iter_addr(vmi) > start)
2613 vma_iter_prev_range(vmi);
2614
2615 error = vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL);
2616 if (error)
2617 goto clear_tree_failed;
2618
2619 /* Point of no return */
2620 mm->locked_vm -= locked_vm;
2621 mm->map_count -= count;
2622 if (unlock)
2623 mmap_write_downgrade(mm);
2624
2625 prev = vma_iter_prev_range(vmi);
2626 next = vma_next(vmi);
2627 if (next)
2628 vma_iter_prev_range(vmi);
2629
2630 /*
2631 * We can free page tables without write-locking mmap_lock because VMAs
2632 * were isolated before we downgraded mmap_lock.
2633 */
2634 mas_set(&mas_detach, 1);
2635 unmap_region(mm, &mas_detach, vma, prev, next, start, end, count,
2636 !unlock);
2637 /* Statistics and freeing VMAs */
2638 mas_set(&mas_detach, 0);
2639 remove_mt(mm, &mas_detach);
2640 validate_mm(mm);
2641 if (unlock)
2642 mmap_read_unlock(mm);
2643
2644 __mt_destroy(&mt_detach);
2645 return 0;
2646
2647clear_tree_failed:
2648userfaultfd_error:
2649munmap_gather_failed:
2650end_split_failed:
2651 mas_set(&mas_detach, 0);
2652 mas_for_each(&mas_detach, next, end)
2653 vma_mark_detached(next, false);
2654
2655 __mt_destroy(&mt_detach);
2656start_split_failed:
2657map_count_exceeded:
2658 validate_mm(mm);
2659 return error;
2660}
2661
2662/*
2663 * do_vmi_munmap() - munmap a given range.
2664 * @vmi: The vma iterator
2665 * @mm: The mm_struct
2666 * @start: The start address to munmap
2667 * @len: The length of the range to munmap
2668 * @uf: The userfaultfd list_head
2669 * @unlock: set to true if the user wants to drop the mmap_lock on success
2670 *
2671 * This function takes a @mas that is either pointing to the previous VMA or set
2672 * to MA_START and sets it up to remove the mapping(s). The @len will be
2673 * aligned and any arch_unmap work will be preformed.
2674 *
2675 * Return: 0 on success and drops the lock if so directed, error and leaves the
2676 * lock held otherwise.
2677 */
2678int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
2679 unsigned long start, size_t len, struct list_head *uf,
2680 bool unlock)
2681{
2682 unsigned long end;
2683 struct vm_area_struct *vma;
2684
2685 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2686 return -EINVAL;
2687
2688 end = start + PAGE_ALIGN(len);
2689 if (end == start)
2690 return -EINVAL;
2691
2692 /* arch_unmap() might do unmaps itself. */
2693 arch_unmap(mm, start, end);
2694
2695 /* Find the first overlapping VMA */
2696 vma = vma_find(vmi, end);
2697 if (!vma) {
2698 if (unlock)
2699 mmap_write_unlock(mm);
2700 return 0;
2701 }
2702
2703 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
2704}
2705
2706/* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2707 * @mm: The mm_struct
2708 * @start: The start address to munmap
2709 * @len: The length to be munmapped.
2710 * @uf: The userfaultfd list_head
2711 *
2712 * Return: 0 on success, error otherwise.
2713 */
2714int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2715 struct list_head *uf)
2716{
2717 VMA_ITERATOR(vmi, mm, start);
2718
2719 return do_vmi_munmap(&vmi, mm, start, len, uf, false);
2720}
2721
2722unsigned long mmap_region(struct file *file, unsigned long addr,
2723 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2724 struct list_head *uf)
2725{
2726 struct mm_struct *mm = current->mm;
2727 struct vm_area_struct *vma = NULL;
2728 struct vm_area_struct *next, *prev, *merge;
2729 pgoff_t pglen = len >> PAGE_SHIFT;
2730 unsigned long charged = 0;
2731 unsigned long end = addr + len;
2732 unsigned long merge_start = addr, merge_end = end;
2733 bool writable_file_mapping = false;
2734 pgoff_t vm_pgoff;
2735 int error;
2736 VMA_ITERATOR(vmi, mm, addr);
2737
2738 /* Check against address space limit. */
2739 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2740 unsigned long nr_pages;
2741
2742 /*
2743 * MAP_FIXED may remove pages of mappings that intersects with
2744 * requested mapping. Account for the pages it would unmap.
2745 */
2746 nr_pages = count_vma_pages_range(mm, addr, end);
2747
2748 if (!may_expand_vm(mm, vm_flags,
2749 (len >> PAGE_SHIFT) - nr_pages))
2750 return -ENOMEM;
2751 }
2752
2753 /* Unmap any existing mapping in the area */
2754 if (do_vmi_munmap(&vmi, mm, addr, len, uf, false))
2755 return -ENOMEM;
2756
2757 /*
2758 * Private writable mapping: check memory availability
2759 */
2760 if (accountable_mapping(file, vm_flags)) {
2761 charged = len >> PAGE_SHIFT;
2762 if (security_vm_enough_memory_mm(mm, charged))
2763 return -ENOMEM;
2764 vm_flags |= VM_ACCOUNT;
2765 }
2766
2767 next = vma_next(&vmi);
2768 prev = vma_prev(&vmi);
2769 if (vm_flags & VM_SPECIAL) {
2770 if (prev)
2771 vma_iter_next_range(&vmi);
2772 goto cannot_expand;
2773 }
2774
2775 /* Attempt to expand an old mapping */
2776 /* Check next */
2777 if (next && next->vm_start == end && !vma_policy(next) &&
2778 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2779 NULL_VM_UFFD_CTX, NULL)) {
2780 merge_end = next->vm_end;
2781 vma = next;
2782 vm_pgoff = next->vm_pgoff - pglen;
2783 }
2784
2785 /* Check prev */
2786 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2787 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2788 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2789 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2790 NULL_VM_UFFD_CTX, NULL))) {
2791 merge_start = prev->vm_start;
2792 vma = prev;
2793 vm_pgoff = prev->vm_pgoff;
2794 } else if (prev) {
2795 vma_iter_next_range(&vmi);
2796 }
2797
2798 /* Actually expand, if possible */
2799 if (vma &&
2800 !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) {
2801 khugepaged_enter_vma(vma, vm_flags);
2802 goto expanded;
2803 }
2804
2805 if (vma == prev)
2806 vma_iter_set(&vmi, addr);
2807cannot_expand:
2808
2809 /*
2810 * Determine the object being mapped and call the appropriate
2811 * specific mapper. the address has already been validated, but
2812 * not unmapped, but the maps are removed from the list.
2813 */
2814 vma = vm_area_alloc(mm);
2815 if (!vma) {
2816 error = -ENOMEM;
2817 goto unacct_error;
2818 }
2819
2820 vma_iter_config(&vmi, addr, end);
2821 vma->vm_start = addr;
2822 vma->vm_end = end;
2823 vm_flags_init(vma, vm_flags);
2824 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2825 vma->vm_pgoff = pgoff;
2826
2827 if (file) {
2828 vma->vm_file = get_file(file);
2829 error = call_mmap(file, vma);
2830 if (error)
2831 goto unmap_and_free_vma;
2832
2833 if (vma_is_shared_maywrite(vma)) {
2834 error = mapping_map_writable(file->f_mapping);
2835 if (error)
2836 goto close_and_free_vma;
2837
2838 writable_file_mapping = true;
2839 }
2840
2841 /*
2842 * Expansion is handled above, merging is handled below.
2843 * Drivers should not alter the address of the VMA.
2844 */
2845 error = -EINVAL;
2846 if (WARN_ON((addr != vma->vm_start)))
2847 goto close_and_free_vma;
2848
2849 vma_iter_config(&vmi, addr, end);
2850 /*
2851 * If vm_flags changed after call_mmap(), we should try merge
2852 * vma again as we may succeed this time.
2853 */
2854 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2855 merge = vma_merge_new_vma(&vmi, prev, vma,
2856 vma->vm_start, vma->vm_end,
2857 vma->vm_pgoff);
2858 if (merge) {
2859 /*
2860 * ->mmap() can change vma->vm_file and fput
2861 * the original file. So fput the vma->vm_file
2862 * here or we would add an extra fput for file
2863 * and cause general protection fault
2864 * ultimately.
2865 */
2866 fput(vma->vm_file);
2867 vm_area_free(vma);
2868 vma = merge;
2869 /* Update vm_flags to pick up the change. */
2870 vm_flags = vma->vm_flags;
2871 goto unmap_writable;
2872 }
2873 }
2874
2875 vm_flags = vma->vm_flags;
2876 } else if (vm_flags & VM_SHARED) {
2877 error = shmem_zero_setup(vma);
2878 if (error)
2879 goto free_vma;
2880 } else {
2881 vma_set_anonymous(vma);
2882 }
2883
2884 if (map_deny_write_exec(vma, vma->vm_flags)) {
2885 error = -EACCES;
2886 goto close_and_free_vma;
2887 }
2888
2889 /* Allow architectures to sanity-check the vm_flags */
2890 error = -EINVAL;
2891 if (!arch_validate_flags(vma->vm_flags))
2892 goto close_and_free_vma;
2893
2894 error = -ENOMEM;
2895 if (vma_iter_prealloc(&vmi, vma))
2896 goto close_and_free_vma;
2897
2898 /* Lock the VMA since it is modified after insertion into VMA tree */
2899 vma_start_write(vma);
2900 vma_iter_store(&vmi, vma);
2901 mm->map_count++;
2902 if (vma->vm_file) {
2903 i_mmap_lock_write(vma->vm_file->f_mapping);
2904 if (vma_is_shared_maywrite(vma))
2905 mapping_allow_writable(vma->vm_file->f_mapping);
2906
2907 flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2908 vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2909 flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2910 i_mmap_unlock_write(vma->vm_file->f_mapping);
2911 }
2912
2913 /*
2914 * vma_merge() calls khugepaged_enter_vma() either, the below
2915 * call covers the non-merge case.
2916 */
2917 khugepaged_enter_vma(vma, vma->vm_flags);
2918
2919 /* Once vma denies write, undo our temporary denial count */
2920unmap_writable:
2921 if (writable_file_mapping)
2922 mapping_unmap_writable(file->f_mapping);
2923 file = vma->vm_file;
2924 ksm_add_vma(vma);
2925expanded:
2926 perf_event_mmap(vma);
2927
2928 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2929 if (vm_flags & VM_LOCKED) {
2930 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2931 is_vm_hugetlb_page(vma) ||
2932 vma == get_gate_vma(current->mm))
2933 vm_flags_clear(vma, VM_LOCKED_MASK);
2934 else
2935 mm->locked_vm += (len >> PAGE_SHIFT);
2936 }
2937
2938 if (file)
2939 uprobe_mmap(vma);
2940
2941 /*
2942 * New (or expanded) vma always get soft dirty status.
2943 * Otherwise user-space soft-dirty page tracker won't
2944 * be able to distinguish situation when vma area unmapped,
2945 * then new mapped in-place (which must be aimed as
2946 * a completely new data area).
2947 */
2948 vm_flags_set(vma, VM_SOFTDIRTY);
2949
2950 vma_set_page_prot(vma);
2951
2952 validate_mm(mm);
2953 return addr;
2954
2955close_and_free_vma:
2956 if (file && vma->vm_ops && vma->vm_ops->close)
2957 vma->vm_ops->close(vma);
2958
2959 if (file || vma->vm_file) {
2960unmap_and_free_vma:
2961 fput(vma->vm_file);
2962 vma->vm_file = NULL;
2963
2964 vma_iter_set(&vmi, vma->vm_end);
2965 /* Undo any partial mapping done by a device driver. */
2966 unmap_region(mm, &vmi.mas, vma, prev, next, vma->vm_start,
2967 vma->vm_end, vma->vm_end, true);
2968 }
2969 if (writable_file_mapping)
2970 mapping_unmap_writable(file->f_mapping);
2971free_vma:
2972 vm_area_free(vma);
2973unacct_error:
2974 if (charged)
2975 vm_unacct_memory(charged);
2976 validate_mm(mm);
2977 return error;
2978}
2979
2980static int __vm_munmap(unsigned long start, size_t len, bool unlock)
2981{
2982 int ret;
2983 struct mm_struct *mm = current->mm;
2984 LIST_HEAD(uf);
2985 VMA_ITERATOR(vmi, mm, start);
2986
2987 if (mmap_write_lock_killable(mm))
2988 return -EINTR;
2989
2990 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, unlock);
2991 if (ret || !unlock)
2992 mmap_write_unlock(mm);
2993
2994 userfaultfd_unmap_complete(mm, &uf);
2995 return ret;
2996}
2997
2998int vm_munmap(unsigned long start, size_t len)
2999{
3000 return __vm_munmap(start, len, false);
3001}
3002EXPORT_SYMBOL(vm_munmap);
3003
3004SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
3005{
3006 addr = untagged_addr(addr);
3007 return __vm_munmap(addr, len, true);
3008}
3009
3010
3011/*
3012 * Emulation of deprecated remap_file_pages() syscall.
3013 */
3014SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
3015 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
3016{
3017
3018 struct mm_struct *mm = current->mm;
3019 struct vm_area_struct *vma;
3020 unsigned long populate = 0;
3021 unsigned long ret = -EINVAL;
3022 struct file *file;
3023
3024 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
3025 current->comm, current->pid);
3026
3027 if (prot)
3028 return ret;
3029 start = start & PAGE_MASK;
3030 size = size & PAGE_MASK;
3031
3032 if (start + size <= start)
3033 return ret;
3034
3035 /* Does pgoff wrap? */
3036 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
3037 return ret;
3038
3039 if (mmap_write_lock_killable(mm))
3040 return -EINTR;
3041
3042 vma = vma_lookup(mm, start);
3043
3044 if (!vma || !(vma->vm_flags & VM_SHARED))
3045 goto out;
3046
3047 if (start + size > vma->vm_end) {
3048 VMA_ITERATOR(vmi, mm, vma->vm_end);
3049 struct vm_area_struct *next, *prev = vma;
3050
3051 for_each_vma_range(vmi, next, start + size) {
3052 /* hole between vmas ? */
3053 if (next->vm_start != prev->vm_end)
3054 goto out;
3055
3056 if (next->vm_file != vma->vm_file)
3057 goto out;
3058
3059 if (next->vm_flags != vma->vm_flags)
3060 goto out;
3061
3062 if (start + size <= next->vm_end)
3063 break;
3064
3065 prev = next;
3066 }
3067
3068 if (!next)
3069 goto out;
3070 }
3071
3072 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3073 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3074 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3075
3076 flags &= MAP_NONBLOCK;
3077 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3078 if (vma->vm_flags & VM_LOCKED)
3079 flags |= MAP_LOCKED;
3080
3081 file = get_file(vma->vm_file);
3082 ret = do_mmap(vma->vm_file, start, size,
3083 prot, flags, 0, pgoff, &populate, NULL);
3084 fput(file);
3085out:
3086 mmap_write_unlock(mm);
3087 if (populate)
3088 mm_populate(ret, populate);
3089 if (!IS_ERR_VALUE(ret))
3090 ret = 0;
3091 return ret;
3092}
3093
3094/*
3095 * do_vma_munmap() - Unmap a full or partial vma.
3096 * @vmi: The vma iterator pointing at the vma
3097 * @vma: The first vma to be munmapped
3098 * @start: the start of the address to unmap
3099 * @end: The end of the address to unmap
3100 * @uf: The userfaultfd list_head
3101 * @unlock: Drop the lock on success
3102 *
3103 * unmaps a VMA mapping when the vma iterator is already in position.
3104 * Does not handle alignment.
3105 *
3106 * Return: 0 on success drops the lock of so directed, error on failure and will
3107 * still hold the lock.
3108 */
3109int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
3110 unsigned long start, unsigned long end, struct list_head *uf,
3111 bool unlock)
3112{
3113 struct mm_struct *mm = vma->vm_mm;
3114
3115 arch_unmap(mm, start, end);
3116 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
3117}
3118
3119/*
3120 * do_brk_flags() - Increase the brk vma if the flags match.
3121 * @vmi: The vma iterator
3122 * @addr: The start address
3123 * @len: The length of the increase
3124 * @vma: The vma,
3125 * @flags: The VMA Flags
3126 *
3127 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
3128 * do not match then create a new anonymous VMA. Eventually we may be able to
3129 * do some brk-specific accounting here.
3130 */
3131static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
3132 unsigned long addr, unsigned long len, unsigned long flags)
3133{
3134 struct mm_struct *mm = current->mm;
3135 struct vma_prepare vp;
3136
3137 /*
3138 * Check against address space limits by the changed size
3139 * Note: This happens *after* clearing old mappings in some code paths.
3140 */
3141 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3142 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3143 return -ENOMEM;
3144
3145 if (mm->map_count > sysctl_max_map_count)
3146 return -ENOMEM;
3147
3148 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3149 return -ENOMEM;
3150
3151 /*
3152 * Expand the existing vma if possible; Note that singular lists do not
3153 * occur after forking, so the expand will only happen on new VMAs.
3154 */
3155 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
3156 can_vma_merge_after(vma, flags, NULL, NULL,
3157 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
3158 vma_iter_config(vmi, vma->vm_start, addr + len);
3159 if (vma_iter_prealloc(vmi, vma))
3160 goto unacct_fail;
3161
3162 vma_start_write(vma);
3163
3164 init_vma_prep(&vp, vma);
3165 vma_prepare(&vp);
3166 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
3167 vma->vm_end = addr + len;
3168 vm_flags_set(vma, VM_SOFTDIRTY);
3169 vma_iter_store(vmi, vma);
3170
3171 vma_complete(&vp, vmi, mm);
3172 khugepaged_enter_vma(vma, flags);
3173 goto out;
3174 }
3175
3176 if (vma)
3177 vma_iter_next_range(vmi);
3178 /* create a vma struct for an anonymous mapping */
3179 vma = vm_area_alloc(mm);
3180 if (!vma)
3181 goto unacct_fail;
3182
3183 vma_set_anonymous(vma);
3184 vma->vm_start = addr;
3185 vma->vm_end = addr + len;
3186 vma->vm_pgoff = addr >> PAGE_SHIFT;
3187 vm_flags_init(vma, flags);
3188 vma->vm_page_prot = vm_get_page_prot(flags);
3189 vma_start_write(vma);
3190 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
3191 goto mas_store_fail;
3192
3193 mm->map_count++;
3194 validate_mm(mm);
3195 ksm_add_vma(vma);
3196out:
3197 perf_event_mmap(vma);
3198 mm->total_vm += len >> PAGE_SHIFT;
3199 mm->data_vm += len >> PAGE_SHIFT;
3200 if (flags & VM_LOCKED)
3201 mm->locked_vm += (len >> PAGE_SHIFT);
3202 vm_flags_set(vma, VM_SOFTDIRTY);
3203 return 0;
3204
3205mas_store_fail:
3206 vm_area_free(vma);
3207unacct_fail:
3208 vm_unacct_memory(len >> PAGE_SHIFT);
3209 return -ENOMEM;
3210}
3211
3212int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3213{
3214 struct mm_struct *mm = current->mm;
3215 struct vm_area_struct *vma = NULL;
3216 unsigned long len;
3217 int ret;
3218 bool populate;
3219 LIST_HEAD(uf);
3220 VMA_ITERATOR(vmi, mm, addr);
3221
3222 len = PAGE_ALIGN(request);
3223 if (len < request)
3224 return -ENOMEM;
3225 if (!len)
3226 return 0;
3227
3228 /* Until we need other flags, refuse anything except VM_EXEC. */
3229 if ((flags & (~VM_EXEC)) != 0)
3230 return -EINVAL;
3231
3232 if (mmap_write_lock_killable(mm))
3233 return -EINTR;
3234
3235 ret = check_brk_limits(addr, len);
3236 if (ret)
3237 goto limits_failed;
3238
3239 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
3240 if (ret)
3241 goto munmap_failed;
3242
3243 vma = vma_prev(&vmi);
3244 ret = do_brk_flags(&vmi, vma, addr, len, flags);
3245 populate = ((mm->def_flags & VM_LOCKED) != 0);
3246 mmap_write_unlock(mm);
3247 userfaultfd_unmap_complete(mm, &uf);
3248 if (populate && !ret)
3249 mm_populate(addr, len);
3250 return ret;
3251
3252munmap_failed:
3253limits_failed:
3254 mmap_write_unlock(mm);
3255 return ret;
3256}
3257EXPORT_SYMBOL(vm_brk_flags);
3258
3259/* Release all mmaps. */
3260void exit_mmap(struct mm_struct *mm)
3261{
3262 struct mmu_gather tlb;
3263 struct vm_area_struct *vma;
3264 unsigned long nr_accounted = 0;
3265 MA_STATE(mas, &mm->mm_mt, 0, 0);
3266 int count = 0;
3267
3268 /* mm's last user has gone, and its about to be pulled down */
3269 mmu_notifier_release(mm);
3270
3271 mmap_read_lock(mm);
3272 arch_exit_mmap(mm);
3273
3274 vma = mas_find(&mas, ULONG_MAX);
3275 if (!vma || unlikely(xa_is_zero(vma))) {
3276 /* Can happen if dup_mmap() received an OOM */
3277 mmap_read_unlock(mm);
3278 mmap_write_lock(mm);
3279 goto destroy;
3280 }
3281
3282 lru_add_drain();
3283 flush_cache_mm(mm);
3284 tlb_gather_mmu_fullmm(&tlb, mm);
3285 /* update_hiwater_rss(mm) here? but nobody should be looking */
3286 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3287 unmap_vmas(&tlb, &mas, vma, 0, ULONG_MAX, ULONG_MAX, false);
3288 mmap_read_unlock(mm);
3289
3290 /*
3291 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3292 * because the memory has been already freed.
3293 */
3294 set_bit(MMF_OOM_SKIP, &mm->flags);
3295 mmap_write_lock(mm);
3296 mt_clear_in_rcu(&mm->mm_mt);
3297 mas_set(&mas, vma->vm_end);
3298 free_pgtables(&tlb, &mas, vma, FIRST_USER_ADDRESS,
3299 USER_PGTABLES_CEILING, true);
3300 tlb_finish_mmu(&tlb);
3301
3302 /*
3303 * Walk the list again, actually closing and freeing it, with preemption
3304 * enabled, without holding any MM locks besides the unreachable
3305 * mmap_write_lock.
3306 */
3307 mas_set(&mas, vma->vm_end);
3308 do {
3309 if (vma->vm_flags & VM_ACCOUNT)
3310 nr_accounted += vma_pages(vma);
3311 remove_vma(vma, true);
3312 count++;
3313 cond_resched();
3314 vma = mas_find(&mas, ULONG_MAX);
3315 } while (vma && likely(!xa_is_zero(vma)));
3316
3317 BUG_ON(count != mm->map_count);
3318
3319 trace_exit_mmap(mm);
3320destroy:
3321 __mt_destroy(&mm->mm_mt);
3322 mmap_write_unlock(mm);
3323 vm_unacct_memory(nr_accounted);
3324}
3325
3326/* Insert vm structure into process list sorted by address
3327 * and into the inode's i_mmap tree. If vm_file is non-NULL
3328 * then i_mmap_rwsem is taken here.
3329 */
3330int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3331{
3332 unsigned long charged = vma_pages(vma);
3333
3334
3335 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3336 return -ENOMEM;
3337
3338 if ((vma->vm_flags & VM_ACCOUNT) &&
3339 security_vm_enough_memory_mm(mm, charged))
3340 return -ENOMEM;
3341
3342 /*
3343 * The vm_pgoff of a purely anonymous vma should be irrelevant
3344 * until its first write fault, when page's anon_vma and index
3345 * are set. But now set the vm_pgoff it will almost certainly
3346 * end up with (unless mremap moves it elsewhere before that
3347 * first wfault), so /proc/pid/maps tells a consistent story.
3348 *
3349 * By setting it to reflect the virtual start address of the
3350 * vma, merges and splits can happen in a seamless way, just
3351 * using the existing file pgoff checks and manipulations.
3352 * Similarly in do_mmap and in do_brk_flags.
3353 */
3354 if (vma_is_anonymous(vma)) {
3355 BUG_ON(vma->anon_vma);
3356 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3357 }
3358
3359 if (vma_link(mm, vma)) {
3360 if (vma->vm_flags & VM_ACCOUNT)
3361 vm_unacct_memory(charged);
3362 return -ENOMEM;
3363 }
3364
3365 return 0;
3366}
3367
3368/*
3369 * Copy the vma structure to a new location in the same mm,
3370 * prior to moving page table entries, to effect an mremap move.
3371 */
3372struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3373 unsigned long addr, unsigned long len, pgoff_t pgoff,
3374 bool *need_rmap_locks)
3375{
3376 struct vm_area_struct *vma = *vmap;
3377 unsigned long vma_start = vma->vm_start;
3378 struct mm_struct *mm = vma->vm_mm;
3379 struct vm_area_struct *new_vma, *prev;
3380 bool faulted_in_anon_vma = true;
3381 VMA_ITERATOR(vmi, mm, addr);
3382
3383 /*
3384 * If anonymous vma has not yet been faulted, update new pgoff
3385 * to match new location, to increase its chance of merging.
3386 */
3387 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3388 pgoff = addr >> PAGE_SHIFT;
3389 faulted_in_anon_vma = false;
3390 }
3391
3392 new_vma = find_vma_prev(mm, addr, &prev);
3393 if (new_vma && new_vma->vm_start < addr + len)
3394 return NULL; /* should never get here */
3395
3396 new_vma = vma_merge_new_vma(&vmi, prev, vma, addr, addr + len, pgoff);
3397 if (new_vma) {
3398 /*
3399 * Source vma may have been merged into new_vma
3400 */
3401 if (unlikely(vma_start >= new_vma->vm_start &&
3402 vma_start < new_vma->vm_end)) {
3403 /*
3404 * The only way we can get a vma_merge with
3405 * self during an mremap is if the vma hasn't
3406 * been faulted in yet and we were allowed to
3407 * reset the dst vma->vm_pgoff to the
3408 * destination address of the mremap to allow
3409 * the merge to happen. mremap must change the
3410 * vm_pgoff linearity between src and dst vmas
3411 * (in turn preventing a vma_merge) to be
3412 * safe. It is only safe to keep the vm_pgoff
3413 * linear if there are no pages mapped yet.
3414 */
3415 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3416 *vmap = vma = new_vma;
3417 }
3418 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3419 } else {
3420 new_vma = vm_area_dup(vma);
3421 if (!new_vma)
3422 goto out;
3423 new_vma->vm_start = addr;
3424 new_vma->vm_end = addr + len;
3425 new_vma->vm_pgoff = pgoff;
3426 if (vma_dup_policy(vma, new_vma))
3427 goto out_free_vma;
3428 if (anon_vma_clone(new_vma, vma))
3429 goto out_free_mempol;
3430 if (new_vma->vm_file)
3431 get_file(new_vma->vm_file);
3432 if (new_vma->vm_ops && new_vma->vm_ops->open)
3433 new_vma->vm_ops->open(new_vma);
3434 if (vma_link(mm, new_vma))
3435 goto out_vma_link;
3436 *need_rmap_locks = false;
3437 }
3438 return new_vma;
3439
3440out_vma_link:
3441 if (new_vma->vm_ops && new_vma->vm_ops->close)
3442 new_vma->vm_ops->close(new_vma);
3443
3444 if (new_vma->vm_file)
3445 fput(new_vma->vm_file);
3446
3447 unlink_anon_vmas(new_vma);
3448out_free_mempol:
3449 mpol_put(vma_policy(new_vma));
3450out_free_vma:
3451 vm_area_free(new_vma);
3452out:
3453 return NULL;
3454}
3455
3456/*
3457 * Return true if the calling process may expand its vm space by the passed
3458 * number of pages
3459 */
3460bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3461{
3462 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3463 return false;
3464
3465 if (is_data_mapping(flags) &&
3466 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3467 /* Workaround for Valgrind */
3468 if (rlimit(RLIMIT_DATA) == 0 &&
3469 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3470 return true;
3471
3472 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3473 current->comm, current->pid,
3474 (mm->data_vm + npages) << PAGE_SHIFT,
3475 rlimit(RLIMIT_DATA),
3476 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3477
3478 if (!ignore_rlimit_data)
3479 return false;
3480 }
3481
3482 return true;
3483}
3484
3485void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3486{
3487 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3488
3489 if (is_exec_mapping(flags))
3490 mm->exec_vm += npages;
3491 else if (is_stack_mapping(flags))
3492 mm->stack_vm += npages;
3493 else if (is_data_mapping(flags))
3494 mm->data_vm += npages;
3495}
3496
3497static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3498
3499/*
3500 * Having a close hook prevents vma merging regardless of flags.
3501 */
3502static void special_mapping_close(struct vm_area_struct *vma)
3503{
3504}
3505
3506static const char *special_mapping_name(struct vm_area_struct *vma)
3507{
3508 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3509}
3510
3511static int special_mapping_mremap(struct vm_area_struct *new_vma)
3512{
3513 struct vm_special_mapping *sm = new_vma->vm_private_data;
3514
3515 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3516 return -EFAULT;
3517
3518 if (sm->mremap)
3519 return sm->mremap(sm, new_vma);
3520
3521 return 0;
3522}
3523
3524static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3525{
3526 /*
3527 * Forbid splitting special mappings - kernel has expectations over
3528 * the number of pages in mapping. Together with VM_DONTEXPAND
3529 * the size of vma should stay the same over the special mapping's
3530 * lifetime.
3531 */
3532 return -EINVAL;
3533}
3534
3535static const struct vm_operations_struct special_mapping_vmops = {
3536 .close = special_mapping_close,
3537 .fault = special_mapping_fault,
3538 .mremap = special_mapping_mremap,
3539 .name = special_mapping_name,
3540 /* vDSO code relies that VVAR can't be accessed remotely */
3541 .access = NULL,
3542 .may_split = special_mapping_split,
3543};
3544
3545static const struct vm_operations_struct legacy_special_mapping_vmops = {
3546 .close = special_mapping_close,
3547 .fault = special_mapping_fault,
3548};
3549
3550static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3551{
3552 struct vm_area_struct *vma = vmf->vma;
3553 pgoff_t pgoff;
3554 struct page **pages;
3555
3556 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3557 pages = vma->vm_private_data;
3558 } else {
3559 struct vm_special_mapping *sm = vma->vm_private_data;
3560
3561 if (sm->fault)
3562 return sm->fault(sm, vmf->vma, vmf);
3563
3564 pages = sm->pages;
3565 }
3566
3567 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3568 pgoff--;
3569
3570 if (*pages) {
3571 struct page *page = *pages;
3572 get_page(page);
3573 vmf->page = page;
3574 return 0;
3575 }
3576
3577 return VM_FAULT_SIGBUS;
3578}
3579
3580static struct vm_area_struct *__install_special_mapping(
3581 struct mm_struct *mm,
3582 unsigned long addr, unsigned long len,
3583 unsigned long vm_flags, void *priv,
3584 const struct vm_operations_struct *ops)
3585{
3586 int ret;
3587 struct vm_area_struct *vma;
3588
3589 vma = vm_area_alloc(mm);
3590 if (unlikely(vma == NULL))
3591 return ERR_PTR(-ENOMEM);
3592
3593 vma->vm_start = addr;
3594 vma->vm_end = addr + len;
3595
3596 vm_flags_init(vma, (vm_flags | mm->def_flags |
3597 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
3598 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3599
3600 vma->vm_ops = ops;
3601 vma->vm_private_data = priv;
3602
3603 ret = insert_vm_struct(mm, vma);
3604 if (ret)
3605 goto out;
3606
3607 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3608
3609 perf_event_mmap(vma);
3610
3611 return vma;
3612
3613out:
3614 vm_area_free(vma);
3615 return ERR_PTR(ret);
3616}
3617
3618bool vma_is_special_mapping(const struct vm_area_struct *vma,
3619 const struct vm_special_mapping *sm)
3620{
3621 return vma->vm_private_data == sm &&
3622 (vma->vm_ops == &special_mapping_vmops ||
3623 vma->vm_ops == &legacy_special_mapping_vmops);
3624}
3625
3626/*
3627 * Called with mm->mmap_lock held for writing.
3628 * Insert a new vma covering the given region, with the given flags.
3629 * Its pages are supplied by the given array of struct page *.
3630 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3631 * The region past the last page supplied will always produce SIGBUS.
3632 * The array pointer and the pages it points to are assumed to stay alive
3633 * for as long as this mapping might exist.
3634 */
3635struct vm_area_struct *_install_special_mapping(
3636 struct mm_struct *mm,
3637 unsigned long addr, unsigned long len,
3638 unsigned long vm_flags, const struct vm_special_mapping *spec)
3639{
3640 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3641 &special_mapping_vmops);
3642}
3643
3644int install_special_mapping(struct mm_struct *mm,
3645 unsigned long addr, unsigned long len,
3646 unsigned long vm_flags, struct page **pages)
3647{
3648 struct vm_area_struct *vma = __install_special_mapping(
3649 mm, addr, len, vm_flags, (void *)pages,
3650 &legacy_special_mapping_vmops);
3651
3652 return PTR_ERR_OR_ZERO(vma);
3653}
3654
3655static DEFINE_MUTEX(mm_all_locks_mutex);
3656
3657static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3658{
3659 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3660 /*
3661 * The LSB of head.next can't change from under us
3662 * because we hold the mm_all_locks_mutex.
3663 */
3664 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3665 /*
3666 * We can safely modify head.next after taking the
3667 * anon_vma->root->rwsem. If some other vma in this mm shares
3668 * the same anon_vma we won't take it again.
3669 *
3670 * No need of atomic instructions here, head.next
3671 * can't change from under us thanks to the
3672 * anon_vma->root->rwsem.
3673 */
3674 if (__test_and_set_bit(0, (unsigned long *)
3675 &anon_vma->root->rb_root.rb_root.rb_node))
3676 BUG();
3677 }
3678}
3679
3680static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3681{
3682 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3683 /*
3684 * AS_MM_ALL_LOCKS can't change from under us because
3685 * we hold the mm_all_locks_mutex.
3686 *
3687 * Operations on ->flags have to be atomic because
3688 * even if AS_MM_ALL_LOCKS is stable thanks to the
3689 * mm_all_locks_mutex, there may be other cpus
3690 * changing other bitflags in parallel to us.
3691 */
3692 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3693 BUG();
3694 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3695 }
3696}
3697
3698/*
3699 * This operation locks against the VM for all pte/vma/mm related
3700 * operations that could ever happen on a certain mm. This includes
3701 * vmtruncate, try_to_unmap, and all page faults.
3702 *
3703 * The caller must take the mmap_lock in write mode before calling
3704 * mm_take_all_locks(). The caller isn't allowed to release the
3705 * mmap_lock until mm_drop_all_locks() returns.
3706 *
3707 * mmap_lock in write mode is required in order to block all operations
3708 * that could modify pagetables and free pages without need of
3709 * altering the vma layout. It's also needed in write mode to avoid new
3710 * anon_vmas to be associated with existing vmas.
3711 *
3712 * A single task can't take more than one mm_take_all_locks() in a row
3713 * or it would deadlock.
3714 *
3715 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3716 * mapping->flags avoid to take the same lock twice, if more than one
3717 * vma in this mm is backed by the same anon_vma or address_space.
3718 *
3719 * We take locks in following order, accordingly to comment at beginning
3720 * of mm/rmap.c:
3721 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3722 * hugetlb mapping);
3723 * - all vmas marked locked
3724 * - all i_mmap_rwsem locks;
3725 * - all anon_vma->rwseml
3726 *
3727 * We can take all locks within these types randomly because the VM code
3728 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3729 * mm_all_locks_mutex.
3730 *
3731 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3732 * that may have to take thousand of locks.
3733 *
3734 * mm_take_all_locks() can fail if it's interrupted by signals.
3735 */
3736int mm_take_all_locks(struct mm_struct *mm)
3737{
3738 struct vm_area_struct *vma;
3739 struct anon_vma_chain *avc;
3740 MA_STATE(mas, &mm->mm_mt, 0, 0);
3741
3742 mmap_assert_write_locked(mm);
3743
3744 mutex_lock(&mm_all_locks_mutex);
3745
3746 /*
3747 * vma_start_write() does not have a complement in mm_drop_all_locks()
3748 * because vma_start_write() is always asymmetrical; it marks a VMA as
3749 * being written to until mmap_write_unlock() or mmap_write_downgrade()
3750 * is reached.
3751 */
3752 mas_for_each(&mas, vma, ULONG_MAX) {
3753 if (signal_pending(current))
3754 goto out_unlock;
3755 vma_start_write(vma);
3756 }
3757
3758 mas_set(&mas, 0);
3759 mas_for_each(&mas, vma, ULONG_MAX) {
3760 if (signal_pending(current))
3761 goto out_unlock;
3762 if (vma->vm_file && vma->vm_file->f_mapping &&
3763 is_vm_hugetlb_page(vma))
3764 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3765 }
3766
3767 mas_set(&mas, 0);
3768 mas_for_each(&mas, vma, ULONG_MAX) {
3769 if (signal_pending(current))
3770 goto out_unlock;
3771 if (vma->vm_file && vma->vm_file->f_mapping &&
3772 !is_vm_hugetlb_page(vma))
3773 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3774 }
3775
3776 mas_set(&mas, 0);
3777 mas_for_each(&mas, vma, ULONG_MAX) {
3778 if (signal_pending(current))
3779 goto out_unlock;
3780 if (vma->anon_vma)
3781 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3782 vm_lock_anon_vma(mm, avc->anon_vma);
3783 }
3784
3785 return 0;
3786
3787out_unlock:
3788 mm_drop_all_locks(mm);
3789 return -EINTR;
3790}
3791
3792static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3793{
3794 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3795 /*
3796 * The LSB of head.next can't change to 0 from under
3797 * us because we hold the mm_all_locks_mutex.
3798 *
3799 * We must however clear the bitflag before unlocking
3800 * the vma so the users using the anon_vma->rb_root will
3801 * never see our bitflag.
3802 *
3803 * No need of atomic instructions here, head.next
3804 * can't change from under us until we release the
3805 * anon_vma->root->rwsem.
3806 */
3807 if (!__test_and_clear_bit(0, (unsigned long *)
3808 &anon_vma->root->rb_root.rb_root.rb_node))
3809 BUG();
3810 anon_vma_unlock_write(anon_vma);
3811 }
3812}
3813
3814static void vm_unlock_mapping(struct address_space *mapping)
3815{
3816 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3817 /*
3818 * AS_MM_ALL_LOCKS can't change to 0 from under us
3819 * because we hold the mm_all_locks_mutex.
3820 */
3821 i_mmap_unlock_write(mapping);
3822 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3823 &mapping->flags))
3824 BUG();
3825 }
3826}
3827
3828/*
3829 * The mmap_lock cannot be released by the caller until
3830 * mm_drop_all_locks() returns.
3831 */
3832void mm_drop_all_locks(struct mm_struct *mm)
3833{
3834 struct vm_area_struct *vma;
3835 struct anon_vma_chain *avc;
3836 MA_STATE(mas, &mm->mm_mt, 0, 0);
3837
3838 mmap_assert_write_locked(mm);
3839 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3840
3841 mas_for_each(&mas, vma, ULONG_MAX) {
3842 if (vma->anon_vma)
3843 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3844 vm_unlock_anon_vma(avc->anon_vma);
3845 if (vma->vm_file && vma->vm_file->f_mapping)
3846 vm_unlock_mapping(vma->vm_file->f_mapping);
3847 }
3848
3849 mutex_unlock(&mm_all_locks_mutex);
3850}
3851
3852/*
3853 * initialise the percpu counter for VM
3854 */
3855void __init mmap_init(void)
3856{
3857 int ret;
3858
3859 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3860 VM_BUG_ON(ret);
3861}
3862
3863/*
3864 * Initialise sysctl_user_reserve_kbytes.
3865 *
3866 * This is intended to prevent a user from starting a single memory hogging
3867 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3868 * mode.
3869 *
3870 * The default value is min(3% of free memory, 128MB)
3871 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3872 */
3873static int init_user_reserve(void)
3874{
3875 unsigned long free_kbytes;
3876
3877 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3878
3879 sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3880 return 0;
3881}
3882subsys_initcall(init_user_reserve);
3883
3884/*
3885 * Initialise sysctl_admin_reserve_kbytes.
3886 *
3887 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3888 * to log in and kill a memory hogging process.
3889 *
3890 * Systems with more than 256MB will reserve 8MB, enough to recover
3891 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3892 * only reserve 3% of free pages by default.
3893 */
3894static int init_admin_reserve(void)
3895{
3896 unsigned long free_kbytes;
3897
3898 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3899
3900 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3901 return 0;
3902}
3903subsys_initcall(init_admin_reserve);
3904
3905/*
3906 * Reinititalise user and admin reserves if memory is added or removed.
3907 *
3908 * The default user reserve max is 128MB, and the default max for the
3909 * admin reserve is 8MB. These are usually, but not always, enough to
3910 * enable recovery from a memory hogging process using login/sshd, a shell,
3911 * and tools like top. It may make sense to increase or even disable the
3912 * reserve depending on the existence of swap or variations in the recovery
3913 * tools. So, the admin may have changed them.
3914 *
3915 * If memory is added and the reserves have been eliminated or increased above
3916 * the default max, then we'll trust the admin.
3917 *
3918 * If memory is removed and there isn't enough free memory, then we
3919 * need to reset the reserves.
3920 *
3921 * Otherwise keep the reserve set by the admin.
3922 */
3923static int reserve_mem_notifier(struct notifier_block *nb,
3924 unsigned long action, void *data)
3925{
3926 unsigned long tmp, free_kbytes;
3927
3928 switch (action) {
3929 case MEM_ONLINE:
3930 /* Default max is 128MB. Leave alone if modified by operator. */
3931 tmp = sysctl_user_reserve_kbytes;
3932 if (0 < tmp && tmp < (1UL << 17))
3933 init_user_reserve();
3934
3935 /* Default max is 8MB. Leave alone if modified by operator. */
3936 tmp = sysctl_admin_reserve_kbytes;
3937 if (0 < tmp && tmp < (1UL << 13))
3938 init_admin_reserve();
3939
3940 break;
3941 case MEM_OFFLINE:
3942 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3943
3944 if (sysctl_user_reserve_kbytes > free_kbytes) {
3945 init_user_reserve();
3946 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3947 sysctl_user_reserve_kbytes);
3948 }
3949
3950 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3951 init_admin_reserve();
3952 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3953 sysctl_admin_reserve_kbytes);
3954 }
3955 break;
3956 default:
3957 break;
3958 }
3959 return NOTIFY_OK;
3960}
3961
3962static int __meminit init_reserve_notifier(void)
3963{
3964 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3965 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3966
3967 return 0;
3968}
3969subsys_initcall(init_reserve_notifier);
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
66const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
67int mmap_rnd_bits_max __ro_after_init = CONFIG_ARCH_MMAP_RND_BITS_MAX;
68int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
69#endif
70#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
71const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
72const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
73int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
74#endif
75
76static bool ignore_rlimit_data;
77core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
78
79static 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
84static 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. */
90void 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 */
107static void __remove_shared_vm_struct(struct vm_area_struct *vma,
108 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 */
122void 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, mapping);
130 i_mmap_unlock_write(mapping);
131 }
132}
133
134/*
135 * Close a vm structure and free it.
136 */
137static 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
151static 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 */
165static 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}
176static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
177 unsigned long addr, unsigned long request, unsigned long flags);
178SYSCALL_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
266success:
267 mmap_write_unlock(mm);
268success_unlocked:
269 userfaultfd_unmap_complete(mm, &uf);
270 if (populate)
271 mm_populate(oldbrk, newbrk - oldbrk);
272 return brk;
273
274out:
275 mm->brk = origbrk;
276 mmap_write_unlock(mm);
277 return origbrk;
278}
279
280#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
281static 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 */
349static inline void
350anon_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
358static inline void
359anon_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
367static 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
384static 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
395static void vma_link_file(struct vm_area_struct *vma)
396{
397 struct file *file = vma->vm_file;
398 struct address_space *mapping;
399
400 if (file) {
401 mapping = file->f_mapping;
402 i_mmap_lock_write(mapping);
403 __vma_link_file(vma, mapping);
404 i_mmap_unlock_write(mapping);
405 }
406}
407
408static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
409{
410 VMA_ITERATOR(vmi, mm, 0);
411
412 vma_iter_config(&vmi, vma->vm_start, vma->vm_end);
413 if (vma_iter_prealloc(&vmi, vma))
414 return -ENOMEM;
415
416 vma_start_write(vma);
417 vma_iter_store(&vmi, vma);
418 vma_link_file(vma);
419 mm->map_count++;
420 validate_mm(mm);
421 return 0;
422}
423
424/*
425 * init_multi_vma_prep() - Initializer for struct vma_prepare
426 * @vp: The vma_prepare struct
427 * @vma: The vma that will be altered once locked
428 * @next: The next vma if it is to be adjusted
429 * @remove: The first vma to be removed
430 * @remove2: The second vma to be removed
431 */
432static inline void init_multi_vma_prep(struct vma_prepare *vp,
433 struct vm_area_struct *vma, struct vm_area_struct *next,
434 struct vm_area_struct *remove, struct vm_area_struct *remove2)
435{
436 memset(vp, 0, sizeof(struct vma_prepare));
437 vp->vma = vma;
438 vp->anon_vma = vma->anon_vma;
439 vp->remove = remove;
440 vp->remove2 = remove2;
441 vp->adj_next = next;
442 if (!vp->anon_vma && next)
443 vp->anon_vma = next->anon_vma;
444
445 vp->file = vma->vm_file;
446 if (vp->file)
447 vp->mapping = vma->vm_file->f_mapping;
448
449}
450
451/*
452 * init_vma_prep() - Initializer wrapper for vma_prepare struct
453 * @vp: The vma_prepare struct
454 * @vma: The vma that will be altered once locked
455 */
456static inline void init_vma_prep(struct vma_prepare *vp,
457 struct vm_area_struct *vma)
458{
459 init_multi_vma_prep(vp, vma, NULL, NULL, NULL);
460}
461
462
463/*
464 * vma_prepare() - Helper function for handling locking VMAs prior to altering
465 * @vp: The initialized vma_prepare struct
466 */
467static inline void vma_prepare(struct vma_prepare *vp)
468{
469 if (vp->file) {
470 uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end);
471
472 if (vp->adj_next)
473 uprobe_munmap(vp->adj_next, vp->adj_next->vm_start,
474 vp->adj_next->vm_end);
475
476 i_mmap_lock_write(vp->mapping);
477 if (vp->insert && vp->insert->vm_file) {
478 /*
479 * Put into interval tree now, so instantiated pages
480 * are visible to arm/parisc __flush_dcache_page
481 * throughout; but we cannot insert into address
482 * space until vma start or end is updated.
483 */
484 __vma_link_file(vp->insert,
485 vp->insert->vm_file->f_mapping);
486 }
487 }
488
489 if (vp->anon_vma) {
490 anon_vma_lock_write(vp->anon_vma);
491 anon_vma_interval_tree_pre_update_vma(vp->vma);
492 if (vp->adj_next)
493 anon_vma_interval_tree_pre_update_vma(vp->adj_next);
494 }
495
496 if (vp->file) {
497 flush_dcache_mmap_lock(vp->mapping);
498 vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap);
499 if (vp->adj_next)
500 vma_interval_tree_remove(vp->adj_next,
501 &vp->mapping->i_mmap);
502 }
503
504}
505
506/*
507 * vma_complete- Helper function for handling the unlocking after altering VMAs,
508 * or for inserting a VMA.
509 *
510 * @vp: The vma_prepare struct
511 * @vmi: The vma iterator
512 * @mm: The mm_struct
513 */
514static inline void vma_complete(struct vma_prepare *vp,
515 struct vma_iterator *vmi, struct mm_struct *mm)
516{
517 if (vp->file) {
518 if (vp->adj_next)
519 vma_interval_tree_insert(vp->adj_next,
520 &vp->mapping->i_mmap);
521 vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap);
522 flush_dcache_mmap_unlock(vp->mapping);
523 }
524
525 if (vp->remove && vp->file) {
526 __remove_shared_vm_struct(vp->remove, vp->mapping);
527 if (vp->remove2)
528 __remove_shared_vm_struct(vp->remove2, vp->mapping);
529 } else if (vp->insert) {
530 /*
531 * split_vma has split insert from vma, and needs
532 * us to insert it before dropping the locks
533 * (it may either follow vma or precede it).
534 */
535 vma_iter_store(vmi, vp->insert);
536 mm->map_count++;
537 }
538
539 if (vp->anon_vma) {
540 anon_vma_interval_tree_post_update_vma(vp->vma);
541 if (vp->adj_next)
542 anon_vma_interval_tree_post_update_vma(vp->adj_next);
543 anon_vma_unlock_write(vp->anon_vma);
544 }
545
546 if (vp->file) {
547 i_mmap_unlock_write(vp->mapping);
548 uprobe_mmap(vp->vma);
549
550 if (vp->adj_next)
551 uprobe_mmap(vp->adj_next);
552 }
553
554 if (vp->remove) {
555again:
556 vma_mark_detached(vp->remove, true);
557 if (vp->file) {
558 uprobe_munmap(vp->remove, vp->remove->vm_start,
559 vp->remove->vm_end);
560 fput(vp->file);
561 }
562 if (vp->remove->anon_vma)
563 anon_vma_merge(vp->vma, vp->remove);
564 mm->map_count--;
565 mpol_put(vma_policy(vp->remove));
566 if (!vp->remove2)
567 WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end);
568 vm_area_free(vp->remove);
569
570 /*
571 * In mprotect's case 6 (see comments on vma_merge),
572 * we are removing both mid and next vmas
573 */
574 if (vp->remove2) {
575 vp->remove = vp->remove2;
576 vp->remove2 = NULL;
577 goto again;
578 }
579 }
580 if (vp->insert && vp->file)
581 uprobe_mmap(vp->insert);
582 validate_mm(mm);
583}
584
585/*
586 * dup_anon_vma() - Helper function to duplicate anon_vma
587 * @dst: The destination VMA
588 * @src: The source VMA
589 * @dup: Pointer to the destination VMA when successful.
590 *
591 * Returns: 0 on success.
592 */
593static inline int dup_anon_vma(struct vm_area_struct *dst,
594 struct vm_area_struct *src, struct vm_area_struct **dup)
595{
596 /*
597 * Easily overlooked: when mprotect shifts the boundary, make sure the
598 * expanding vma has anon_vma set if the shrinking vma had, to cover any
599 * anon pages imported.
600 */
601 if (src->anon_vma && !dst->anon_vma) {
602 int ret;
603
604 vma_assert_write_locked(dst);
605 dst->anon_vma = src->anon_vma;
606 ret = anon_vma_clone(dst, src);
607 if (ret)
608 return ret;
609
610 *dup = dst;
611 }
612
613 return 0;
614}
615
616/*
617 * vma_expand - Expand an existing VMA
618 *
619 * @vmi: The vma iterator
620 * @vma: The vma to expand
621 * @start: The start of the vma
622 * @end: The exclusive end of the vma
623 * @pgoff: The page offset of vma
624 * @next: The current of next vma.
625 *
626 * Expand @vma to @start and @end. Can expand off the start and end. Will
627 * expand over @next if it's different from @vma and @end == @next->vm_end.
628 * Checking if the @vma can expand and merge with @next needs to be handled by
629 * the caller.
630 *
631 * Returns: 0 on success
632 */
633int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
634 unsigned long start, unsigned long end, pgoff_t pgoff,
635 struct vm_area_struct *next)
636{
637 struct vm_area_struct *anon_dup = NULL;
638 bool remove_next = false;
639 struct vma_prepare vp;
640
641 vma_start_write(vma);
642 if (next && (vma != next) && (end == next->vm_end)) {
643 int ret;
644
645 remove_next = true;
646 vma_start_write(next);
647 ret = dup_anon_vma(vma, next, &anon_dup);
648 if (ret)
649 return ret;
650 }
651
652 init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
653 /* Not merging but overwriting any part of next is not handled. */
654 VM_WARN_ON(next && !vp.remove &&
655 next != vma && end > next->vm_start);
656 /* Only handles expanding */
657 VM_WARN_ON(vma->vm_start < start || vma->vm_end > end);
658
659 /* Note: vma iterator must be pointing to 'start' */
660 vma_iter_config(vmi, start, end);
661 if (vma_iter_prealloc(vmi, vma))
662 goto nomem;
663
664 vma_prepare(&vp);
665 vma_adjust_trans_huge(vma, start, end, 0);
666 vma_set_range(vma, start, end, pgoff);
667 vma_iter_store(vmi, vma);
668
669 vma_complete(&vp, vmi, vma->vm_mm);
670 return 0;
671
672nomem:
673 if (anon_dup)
674 unlink_anon_vmas(anon_dup);
675 return -ENOMEM;
676}
677
678/*
679 * vma_shrink() - Reduce an existing VMAs memory area
680 * @vmi: The vma iterator
681 * @vma: The VMA to modify
682 * @start: The new start
683 * @end: The new end
684 *
685 * Returns: 0 on success, -ENOMEM otherwise
686 */
687int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
688 unsigned long start, unsigned long end, pgoff_t pgoff)
689{
690 struct vma_prepare vp;
691
692 WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
693
694 if (vma->vm_start < start)
695 vma_iter_config(vmi, vma->vm_start, start);
696 else
697 vma_iter_config(vmi, end, vma->vm_end);
698
699 if (vma_iter_prealloc(vmi, NULL))
700 return -ENOMEM;
701
702 vma_start_write(vma);
703
704 init_vma_prep(&vp, vma);
705 vma_prepare(&vp);
706 vma_adjust_trans_huge(vma, start, end, 0);
707
708 vma_iter_clear(vmi);
709 vma_set_range(vma, start, end, pgoff);
710 vma_complete(&vp, vmi, vma->vm_mm);
711 return 0;
712}
713
714/*
715 * If the vma has a ->close operation then the driver probably needs to release
716 * per-vma resources, so we don't attempt to merge those if the caller indicates
717 * the current vma may be removed as part of the merge.
718 */
719static inline bool is_mergeable_vma(struct vm_area_struct *vma,
720 struct file *file, unsigned long vm_flags,
721 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
722 struct anon_vma_name *anon_name, bool may_remove_vma)
723{
724 /*
725 * VM_SOFTDIRTY should not prevent from VMA merging, if we
726 * match the flags but dirty bit -- the caller should mark
727 * merged VMA as dirty. If dirty bit won't be excluded from
728 * comparison, we increase pressure on the memory system forcing
729 * the kernel to generate new VMAs when old one could be
730 * extended instead.
731 */
732 if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
733 return false;
734 if (vma->vm_file != file)
735 return false;
736 if (may_remove_vma && vma->vm_ops && vma->vm_ops->close)
737 return false;
738 if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
739 return false;
740 if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
741 return false;
742 return true;
743}
744
745static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1,
746 struct anon_vma *anon_vma2, struct vm_area_struct *vma)
747{
748 /*
749 * The list_is_singular() test is to avoid merging VMA cloned from
750 * parents. This can improve scalability caused by anon_vma lock.
751 */
752 if ((!anon_vma1 || !anon_vma2) && (!vma ||
753 list_is_singular(&vma->anon_vma_chain)))
754 return true;
755 return anon_vma1 == anon_vma2;
756}
757
758/*
759 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
760 * in front of (at a lower virtual address and file offset than) the vma.
761 *
762 * We cannot merge two vmas if they have differently assigned (non-NULL)
763 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
764 *
765 * We don't check here for the merged mmap wrapping around the end of pagecache
766 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
767 * wrap, nor mmaps which cover the final page at index -1UL.
768 *
769 * We assume the vma may be removed as part of the merge.
770 */
771static bool
772can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
773 struct anon_vma *anon_vma, struct file *file,
774 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
775 struct anon_vma_name *anon_name)
776{
777 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, true) &&
778 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
779 if (vma->vm_pgoff == vm_pgoff)
780 return true;
781 }
782 return false;
783}
784
785/*
786 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
787 * beyond (at a higher virtual address and file offset than) the vma.
788 *
789 * We cannot merge two vmas if they have differently assigned (non-NULL)
790 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
791 *
792 * We assume that vma is not removed as part of the merge.
793 */
794static bool
795can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
796 struct anon_vma *anon_vma, struct file *file,
797 pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
798 struct anon_vma_name *anon_name)
799{
800 if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, false) &&
801 is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
802 pgoff_t vm_pglen;
803 vm_pglen = vma_pages(vma);
804 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
805 return true;
806 }
807 return false;
808}
809
810/*
811 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
812 * figure out whether that can be merged with its predecessor or its
813 * successor. Or both (it neatly fills a hole).
814 *
815 * In most cases - when called for mmap, brk or mremap - [addr,end) is
816 * certain not to be mapped by the time vma_merge is called; but when
817 * called for mprotect, it is certain to be already mapped (either at
818 * an offset within prev, or at the start of next), and the flags of
819 * this area are about to be changed to vm_flags - and the no-change
820 * case has already been eliminated.
821 *
822 * The following mprotect cases have to be considered, where **** is
823 * the area passed down from mprotect_fixup, never extending beyond one
824 * vma, PPPP is the previous vma, CCCC is a concurrent vma that starts
825 * at the same address as **** and is of the same or larger span, and
826 * NNNN the next vma after ****:
827 *
828 * **** **** ****
829 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPCCCCCC
830 * cannot merge might become might become
831 * PPNNNNNNNNNN PPPPPPPPPPCC
832 * mmap, brk or case 4 below case 5 below
833 * mremap move:
834 * **** ****
835 * PPPP NNNN PPPPCCCCNNNN
836 * might become might become
837 * PPPPPPPPPPPP 1 or PPPPPPPPPPPP 6 or
838 * PPPPPPPPNNNN 2 or PPPPPPPPNNNN 7 or
839 * PPPPNNNNNNNN 3 PPPPNNNNNNNN 8
840 *
841 * It is important for case 8 that the vma CCCC overlapping the
842 * region **** is never going to extended over NNNN. Instead NNNN must
843 * be extended in region **** and CCCC must be removed. This way in
844 * all cases where vma_merge succeeds, the moment vma_merge drops the
845 * rmap_locks, the properties of the merged vma will be already
846 * correct for the whole merged range. Some of those properties like
847 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
848 * be correct for the whole merged range immediately after the
849 * rmap_locks are released. Otherwise if NNNN would be removed and
850 * CCCC would be extended over the NNNN range, remove_migration_ptes
851 * or other rmap walkers (if working on addresses beyond the "end"
852 * parameter) may establish ptes with the wrong permissions of CCCC
853 * instead of the right permissions of NNNN.
854 *
855 * In the code below:
856 * PPPP is represented by *prev
857 * CCCC is represented by *curr or not represented at all (NULL)
858 * NNNN is represented by *next or not represented at all (NULL)
859 * **** is not represented - it will be merged and the vma containing the
860 * area is returned, or the function will return NULL
861 */
862static struct vm_area_struct
863*vma_merge(struct vma_iterator *vmi, struct vm_area_struct *prev,
864 struct vm_area_struct *src, unsigned long addr, unsigned long end,
865 unsigned long vm_flags, pgoff_t pgoff, struct mempolicy *policy,
866 struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
867 struct anon_vma_name *anon_name)
868{
869 struct mm_struct *mm = src->vm_mm;
870 struct anon_vma *anon_vma = src->anon_vma;
871 struct file *file = src->vm_file;
872 struct vm_area_struct *curr, *next, *res;
873 struct vm_area_struct *vma, *adjust, *remove, *remove2;
874 struct vm_area_struct *anon_dup = NULL;
875 struct vma_prepare vp;
876 pgoff_t vma_pgoff;
877 int err = 0;
878 bool merge_prev = false;
879 bool merge_next = false;
880 bool vma_expanded = false;
881 unsigned long vma_start = addr;
882 unsigned long vma_end = end;
883 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
884 long adj_start = 0;
885
886 /*
887 * We later require that vma->vm_flags == vm_flags,
888 * so this tests vma->vm_flags & VM_SPECIAL, too.
889 */
890 if (vm_flags & VM_SPECIAL)
891 return NULL;
892
893 /* Does the input range span an existing VMA? (cases 5 - 8) */
894 curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end);
895
896 if (!curr || /* cases 1 - 4 */
897 end == curr->vm_end) /* cases 6 - 8, adjacent VMA */
898 next = vma_lookup(mm, end);
899 else
900 next = NULL; /* case 5 */
901
902 if (prev) {
903 vma_start = prev->vm_start;
904 vma_pgoff = prev->vm_pgoff;
905
906 /* Can we merge the predecessor? */
907 if (addr == prev->vm_end && mpol_equal(vma_policy(prev), policy)
908 && can_vma_merge_after(prev, vm_flags, anon_vma, file,
909 pgoff, vm_userfaultfd_ctx, anon_name)) {
910 merge_prev = true;
911 vma_prev(vmi);
912 }
913 }
914
915 /* Can we merge the successor? */
916 if (next && mpol_equal(policy, vma_policy(next)) &&
917 can_vma_merge_before(next, vm_flags, anon_vma, file, pgoff+pglen,
918 vm_userfaultfd_ctx, anon_name)) {
919 merge_next = true;
920 }
921
922 /* Verify some invariant that must be enforced by the caller. */
923 VM_WARN_ON(prev && addr <= prev->vm_start);
924 VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
925 VM_WARN_ON(addr >= end);
926
927 if (!merge_prev && !merge_next)
928 return NULL; /* Not mergeable. */
929
930 if (merge_prev)
931 vma_start_write(prev);
932
933 res = vma = prev;
934 remove = remove2 = adjust = NULL;
935
936 /* Can we merge both the predecessor and the successor? */
937 if (merge_prev && merge_next &&
938 is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) {
939 vma_start_write(next);
940 remove = next; /* case 1 */
941 vma_end = next->vm_end;
942 err = dup_anon_vma(prev, next, &anon_dup);
943 if (curr) { /* case 6 */
944 vma_start_write(curr);
945 remove = curr;
946 remove2 = next;
947 /*
948 * Note that the dup_anon_vma below cannot overwrite err
949 * since the first caller would do nothing unless next
950 * has an anon_vma.
951 */
952 if (!next->anon_vma)
953 err = dup_anon_vma(prev, curr, &anon_dup);
954 }
955 } else if (merge_prev) { /* case 2 */
956 if (curr) {
957 vma_start_write(curr);
958 if (end == curr->vm_end) { /* case 7 */
959 /*
960 * can_vma_merge_after() assumed we would not be
961 * removing prev vma, so it skipped the check
962 * for vm_ops->close, but we are removing curr
963 */
964 if (curr->vm_ops && curr->vm_ops->close)
965 err = -EINVAL;
966 remove = curr;
967 } else { /* case 5 */
968 adjust = curr;
969 adj_start = (end - curr->vm_start);
970 }
971 if (!err)
972 err = dup_anon_vma(prev, curr, &anon_dup);
973 }
974 } else { /* merge_next */
975 vma_start_write(next);
976 res = next;
977 if (prev && addr < prev->vm_end) { /* case 4 */
978 vma_start_write(prev);
979 vma_end = addr;
980 adjust = next;
981 adj_start = -(prev->vm_end - addr);
982 err = dup_anon_vma(next, prev, &anon_dup);
983 } else {
984 /*
985 * Note that cases 3 and 8 are the ONLY ones where prev
986 * is permitted to be (but is not necessarily) NULL.
987 */
988 vma = next; /* case 3 */
989 vma_start = addr;
990 vma_end = next->vm_end;
991 vma_pgoff = next->vm_pgoff - pglen;
992 if (curr) { /* case 8 */
993 vma_pgoff = curr->vm_pgoff;
994 vma_start_write(curr);
995 remove = curr;
996 err = dup_anon_vma(next, curr, &anon_dup);
997 }
998 }
999 }
1000
1001 /* Error in anon_vma clone. */
1002 if (err)
1003 goto anon_vma_fail;
1004
1005 if (vma_start < vma->vm_start || vma_end > vma->vm_end)
1006 vma_expanded = true;
1007
1008 if (vma_expanded) {
1009 vma_iter_config(vmi, vma_start, vma_end);
1010 } else {
1011 vma_iter_config(vmi, adjust->vm_start + adj_start,
1012 adjust->vm_end);
1013 }
1014
1015 if (vma_iter_prealloc(vmi, vma))
1016 goto prealloc_fail;
1017
1018 init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
1019 VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
1020 vp.anon_vma != adjust->anon_vma);
1021
1022 vma_prepare(&vp);
1023 vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1024 vma_set_range(vma, vma_start, vma_end, vma_pgoff);
1025
1026 if (vma_expanded)
1027 vma_iter_store(vmi, vma);
1028
1029 if (adj_start) {
1030 adjust->vm_start += adj_start;
1031 adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1032 if (adj_start < 0) {
1033 WARN_ON(vma_expanded);
1034 vma_iter_store(vmi, next);
1035 }
1036 }
1037
1038 vma_complete(&vp, vmi, mm);
1039 khugepaged_enter_vma(res, vm_flags);
1040 return res;
1041
1042prealloc_fail:
1043 if (anon_dup)
1044 unlink_anon_vmas(anon_dup);
1045
1046anon_vma_fail:
1047 vma_iter_set(vmi, addr);
1048 vma_iter_load(vmi);
1049 return NULL;
1050}
1051
1052/*
1053 * Rough compatibility check to quickly see if it's even worth looking
1054 * at sharing an anon_vma.
1055 *
1056 * They need to have the same vm_file, and the flags can only differ
1057 * in things that mprotect may change.
1058 *
1059 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1060 * we can merge the two vma's. For example, we refuse to merge a vma if
1061 * there is a vm_ops->close() function, because that indicates that the
1062 * driver is doing some kind of reference counting. But that doesn't
1063 * really matter for the anon_vma sharing case.
1064 */
1065static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1066{
1067 return a->vm_end == b->vm_start &&
1068 mpol_equal(vma_policy(a), vma_policy(b)) &&
1069 a->vm_file == b->vm_file &&
1070 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1071 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1072}
1073
1074/*
1075 * Do some basic sanity checking to see if we can re-use the anon_vma
1076 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1077 * the same as 'old', the other will be the new one that is trying
1078 * to share the anon_vma.
1079 *
1080 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1081 * the anon_vma of 'old' is concurrently in the process of being set up
1082 * by another page fault trying to merge _that_. But that's ok: if it
1083 * is being set up, that automatically means that it will be a singleton
1084 * acceptable for merging, so we can do all of this optimistically. But
1085 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1086 *
1087 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1088 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1089 * is to return an anon_vma that is "complex" due to having gone through
1090 * a fork).
1091 *
1092 * We also make sure that the two vma's are compatible (adjacent,
1093 * and with the same memory policies). That's all stable, even with just
1094 * a read lock on the mmap_lock.
1095 */
1096static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1097{
1098 if (anon_vma_compatible(a, b)) {
1099 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1100
1101 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1102 return anon_vma;
1103 }
1104 return NULL;
1105}
1106
1107/*
1108 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1109 * neighbouring vmas for a suitable anon_vma, before it goes off
1110 * to allocate a new anon_vma. It checks because a repetitive
1111 * sequence of mprotects and faults may otherwise lead to distinct
1112 * anon_vmas being allocated, preventing vma merge in subsequent
1113 * mprotect.
1114 */
1115struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1116{
1117 MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1118 struct anon_vma *anon_vma = NULL;
1119 struct vm_area_struct *prev, *next;
1120
1121 /* Try next first. */
1122 next = mas_walk(&mas);
1123 if (next) {
1124 anon_vma = reusable_anon_vma(next, vma, next);
1125 if (anon_vma)
1126 return anon_vma;
1127 }
1128
1129 prev = mas_prev(&mas, 0);
1130 VM_BUG_ON_VMA(prev != vma, vma);
1131 prev = mas_prev(&mas, 0);
1132 /* Try prev next. */
1133 if (prev)
1134 anon_vma = reusable_anon_vma(prev, prev, vma);
1135
1136 /*
1137 * We might reach here with anon_vma == NULL if we can't find
1138 * any reusable anon_vma.
1139 * There's no absolute need to look only at touching neighbours:
1140 * we could search further afield for "compatible" anon_vmas.
1141 * But it would probably just be a waste of time searching,
1142 * or lead to too many vmas hanging off the same anon_vma.
1143 * We're trying to allow mprotect remerging later on,
1144 * not trying to minimize memory used for anon_vmas.
1145 */
1146 return anon_vma;
1147}
1148
1149/*
1150 * If a hint addr is less than mmap_min_addr change hint to be as
1151 * low as possible but still greater than mmap_min_addr
1152 */
1153static inline unsigned long round_hint_to_min(unsigned long hint)
1154{
1155 hint &= PAGE_MASK;
1156 if (((void *)hint != NULL) &&
1157 (hint < mmap_min_addr))
1158 return PAGE_ALIGN(mmap_min_addr);
1159 return hint;
1160}
1161
1162bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
1163 unsigned long bytes)
1164{
1165 unsigned long locked_pages, limit_pages;
1166
1167 if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK))
1168 return true;
1169
1170 locked_pages = bytes >> PAGE_SHIFT;
1171 locked_pages += mm->locked_vm;
1172
1173 limit_pages = rlimit(RLIMIT_MEMLOCK);
1174 limit_pages >>= PAGE_SHIFT;
1175
1176 return locked_pages <= limit_pages;
1177}
1178
1179static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1180{
1181 if (S_ISREG(inode->i_mode))
1182 return MAX_LFS_FILESIZE;
1183
1184 if (S_ISBLK(inode->i_mode))
1185 return MAX_LFS_FILESIZE;
1186
1187 if (S_ISSOCK(inode->i_mode))
1188 return MAX_LFS_FILESIZE;
1189
1190 /* Special "we do even unsigned file positions" case */
1191 if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1192 return 0;
1193
1194 /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1195 return ULONG_MAX;
1196}
1197
1198static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1199 unsigned long pgoff, unsigned long len)
1200{
1201 u64 maxsize = file_mmap_size_max(file, inode);
1202
1203 if (maxsize && len > maxsize)
1204 return false;
1205 maxsize -= len;
1206 if (pgoff > maxsize >> PAGE_SHIFT)
1207 return false;
1208 return true;
1209}
1210
1211/*
1212 * The caller must write-lock current->mm->mmap_lock.
1213 */
1214unsigned long do_mmap(struct file *file, unsigned long addr,
1215 unsigned long len, unsigned long prot,
1216 unsigned long flags, vm_flags_t vm_flags,
1217 unsigned long pgoff, unsigned long *populate,
1218 struct list_head *uf)
1219{
1220 struct mm_struct *mm = current->mm;
1221 int pkey = 0;
1222
1223 *populate = 0;
1224
1225 if (!len)
1226 return -EINVAL;
1227
1228 /*
1229 * Does the application expect PROT_READ to imply PROT_EXEC?
1230 *
1231 * (the exception is when the underlying filesystem is noexec
1232 * mounted, in which case we don't add PROT_EXEC.)
1233 */
1234 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1235 if (!(file && path_noexec(&file->f_path)))
1236 prot |= PROT_EXEC;
1237
1238 /* force arch specific MAP_FIXED handling in get_unmapped_area */
1239 if (flags & MAP_FIXED_NOREPLACE)
1240 flags |= MAP_FIXED;
1241
1242 if (!(flags & MAP_FIXED))
1243 addr = round_hint_to_min(addr);
1244
1245 /* Careful about overflows.. */
1246 len = PAGE_ALIGN(len);
1247 if (!len)
1248 return -ENOMEM;
1249
1250 /* offset overflow? */
1251 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1252 return -EOVERFLOW;
1253
1254 /* Too many mappings? */
1255 if (mm->map_count > sysctl_max_map_count)
1256 return -ENOMEM;
1257
1258 /* Obtain the address to map to. we verify (or select) it and ensure
1259 * that it represents a valid section of the address space.
1260 */
1261 addr = get_unmapped_area(file, addr, len, pgoff, flags);
1262 if (IS_ERR_VALUE(addr))
1263 return addr;
1264
1265 if (flags & MAP_FIXED_NOREPLACE) {
1266 if (find_vma_intersection(mm, addr, addr + len))
1267 return -EEXIST;
1268 }
1269
1270 if (prot == PROT_EXEC) {
1271 pkey = execute_only_pkey(mm);
1272 if (pkey < 0)
1273 pkey = 0;
1274 }
1275
1276 /* Do simple checking here so the lower-level routines won't have
1277 * to. we assume access permissions have been handled by the open
1278 * of the memory object, so we don't do any here.
1279 */
1280 vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1281 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1282
1283 if (flags & MAP_LOCKED)
1284 if (!can_do_mlock())
1285 return -EPERM;
1286
1287 if (!mlock_future_ok(mm, vm_flags, len))
1288 return -EAGAIN;
1289
1290 if (file) {
1291 struct inode *inode = file_inode(file);
1292 unsigned long flags_mask;
1293
1294 if (!file_mmap_ok(file, inode, pgoff, len))
1295 return -EOVERFLOW;
1296
1297 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1298
1299 switch (flags & MAP_TYPE) {
1300 case MAP_SHARED:
1301 /*
1302 * Force use of MAP_SHARED_VALIDATE with non-legacy
1303 * flags. E.g. MAP_SYNC is dangerous to use with
1304 * MAP_SHARED as you don't know which consistency model
1305 * you will get. We silently ignore unsupported flags
1306 * with MAP_SHARED to preserve backward compatibility.
1307 */
1308 flags &= LEGACY_MAP_MASK;
1309 fallthrough;
1310 case MAP_SHARED_VALIDATE:
1311 if (flags & ~flags_mask)
1312 return -EOPNOTSUPP;
1313 if (prot & PROT_WRITE) {
1314 if (!(file->f_mode & FMODE_WRITE))
1315 return -EACCES;
1316 if (IS_SWAPFILE(file->f_mapping->host))
1317 return -ETXTBSY;
1318 }
1319
1320 /*
1321 * Make sure we don't allow writing to an append-only
1322 * file..
1323 */
1324 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1325 return -EACCES;
1326
1327 vm_flags |= VM_SHARED | VM_MAYSHARE;
1328 if (!(file->f_mode & FMODE_WRITE))
1329 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1330 fallthrough;
1331 case MAP_PRIVATE:
1332 if (!(file->f_mode & FMODE_READ))
1333 return -EACCES;
1334 if (path_noexec(&file->f_path)) {
1335 if (vm_flags & VM_EXEC)
1336 return -EPERM;
1337 vm_flags &= ~VM_MAYEXEC;
1338 }
1339
1340 if (!file->f_op->mmap)
1341 return -ENODEV;
1342 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1343 return -EINVAL;
1344 break;
1345
1346 default:
1347 return -EINVAL;
1348 }
1349 } else {
1350 switch (flags & MAP_TYPE) {
1351 case MAP_SHARED:
1352 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1353 return -EINVAL;
1354 /*
1355 * Ignore pgoff.
1356 */
1357 pgoff = 0;
1358 vm_flags |= VM_SHARED | VM_MAYSHARE;
1359 break;
1360 case MAP_PRIVATE:
1361 /*
1362 * Set pgoff according to addr for anon_vma.
1363 */
1364 pgoff = addr >> PAGE_SHIFT;
1365 break;
1366 default:
1367 return -EINVAL;
1368 }
1369 }
1370
1371 /*
1372 * Set 'VM_NORESERVE' if we should not account for the
1373 * memory use of this mapping.
1374 */
1375 if (flags & MAP_NORESERVE) {
1376 /* We honor MAP_NORESERVE if allowed to overcommit */
1377 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1378 vm_flags |= VM_NORESERVE;
1379
1380 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1381 if (file && is_file_hugepages(file))
1382 vm_flags |= VM_NORESERVE;
1383 }
1384
1385 addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1386 if (!IS_ERR_VALUE(addr) &&
1387 ((vm_flags & VM_LOCKED) ||
1388 (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1389 *populate = len;
1390 return addr;
1391}
1392
1393unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1394 unsigned long prot, unsigned long flags,
1395 unsigned long fd, unsigned long pgoff)
1396{
1397 struct file *file = NULL;
1398 unsigned long retval;
1399
1400 if (!(flags & MAP_ANONYMOUS)) {
1401 audit_mmap_fd(fd, flags);
1402 file = fget(fd);
1403 if (!file)
1404 return -EBADF;
1405 if (is_file_hugepages(file)) {
1406 len = ALIGN(len, huge_page_size(hstate_file(file)));
1407 } else if (unlikely(flags & MAP_HUGETLB)) {
1408 retval = -EINVAL;
1409 goto out_fput;
1410 }
1411 } else if (flags & MAP_HUGETLB) {
1412 struct hstate *hs;
1413
1414 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1415 if (!hs)
1416 return -EINVAL;
1417
1418 len = ALIGN(len, huge_page_size(hs));
1419 /*
1420 * VM_NORESERVE is used because the reservations will be
1421 * taken when vm_ops->mmap() is called
1422 */
1423 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1424 VM_NORESERVE,
1425 HUGETLB_ANONHUGE_INODE,
1426 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1427 if (IS_ERR(file))
1428 return PTR_ERR(file);
1429 }
1430
1431 retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1432out_fput:
1433 if (file)
1434 fput(file);
1435 return retval;
1436}
1437
1438SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1439 unsigned long, prot, unsigned long, flags,
1440 unsigned long, fd, unsigned long, pgoff)
1441{
1442 return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1443}
1444
1445#ifdef __ARCH_WANT_SYS_OLD_MMAP
1446struct mmap_arg_struct {
1447 unsigned long addr;
1448 unsigned long len;
1449 unsigned long prot;
1450 unsigned long flags;
1451 unsigned long fd;
1452 unsigned long offset;
1453};
1454
1455SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1456{
1457 struct mmap_arg_struct a;
1458
1459 if (copy_from_user(&a, arg, sizeof(a)))
1460 return -EFAULT;
1461 if (offset_in_page(a.offset))
1462 return -EINVAL;
1463
1464 return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1465 a.offset >> PAGE_SHIFT);
1466}
1467#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1468
1469static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops)
1470{
1471 return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite);
1472}
1473
1474static bool vma_is_shared_writable(struct vm_area_struct *vma)
1475{
1476 return (vma->vm_flags & (VM_WRITE | VM_SHARED)) ==
1477 (VM_WRITE | VM_SHARED);
1478}
1479
1480static bool vma_fs_can_writeback(struct vm_area_struct *vma)
1481{
1482 /* No managed pages to writeback. */
1483 if (vma->vm_flags & VM_PFNMAP)
1484 return false;
1485
1486 return vma->vm_file && vma->vm_file->f_mapping &&
1487 mapping_can_writeback(vma->vm_file->f_mapping);
1488}
1489
1490/*
1491 * Does this VMA require the underlying folios to have their dirty state
1492 * tracked?
1493 */
1494bool vma_needs_dirty_tracking(struct vm_area_struct *vma)
1495{
1496 /* Only shared, writable VMAs require dirty tracking. */
1497 if (!vma_is_shared_writable(vma))
1498 return false;
1499
1500 /* Does the filesystem need to be notified? */
1501 if (vm_ops_needs_writenotify(vma->vm_ops))
1502 return true;
1503
1504 /*
1505 * Even if the filesystem doesn't indicate a need for writenotify, if it
1506 * can writeback, dirty tracking is still required.
1507 */
1508 return vma_fs_can_writeback(vma);
1509}
1510
1511/*
1512 * Some shared mappings will want the pages marked read-only
1513 * to track write events. If so, we'll downgrade vm_page_prot
1514 * to the private version (using protection_map[] without the
1515 * VM_SHARED bit).
1516 */
1517int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1518{
1519 /* If it was private or non-writable, the write bit is already clear */
1520 if (!vma_is_shared_writable(vma))
1521 return 0;
1522
1523 /* The backer wishes to know when pages are first written to? */
1524 if (vm_ops_needs_writenotify(vma->vm_ops))
1525 return 1;
1526
1527 /* The open routine did something to the protections that pgprot_modify
1528 * won't preserve? */
1529 if (pgprot_val(vm_page_prot) !=
1530 pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags)))
1531 return 0;
1532
1533 /*
1534 * Do we need to track softdirty? hugetlb does not support softdirty
1535 * tracking yet.
1536 */
1537 if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1538 return 1;
1539
1540 /* Do we need write faults for uffd-wp tracking? */
1541 if (userfaultfd_wp(vma))
1542 return 1;
1543
1544 /* Can the mapping track the dirty pages? */
1545 return vma_fs_can_writeback(vma);
1546}
1547
1548/*
1549 * We account for memory if it's a private writeable mapping,
1550 * not hugepages and VM_NORESERVE wasn't set.
1551 */
1552static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1553{
1554 /*
1555 * hugetlb has its own accounting separate from the core VM
1556 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1557 */
1558 if (file && is_file_hugepages(file))
1559 return 0;
1560
1561 return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1562}
1563
1564/**
1565 * unmapped_area() - Find an area between the low_limit and the high_limit with
1566 * the correct alignment and offset, all from @info. Note: current->mm is used
1567 * for the search.
1568 *
1569 * @info: The unmapped area information including the range [low_limit -
1570 * high_limit), the alignment offset and mask.
1571 *
1572 * Return: A memory address or -ENOMEM.
1573 */
1574static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1575{
1576 unsigned long length, gap;
1577 unsigned long low_limit, high_limit;
1578 struct vm_area_struct *tmp;
1579
1580 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1581
1582 /* Adjust search length to account for worst case alignment overhead */
1583 length = info->length + info->align_mask;
1584 if (length < info->length)
1585 return -ENOMEM;
1586
1587 low_limit = info->low_limit;
1588 if (low_limit < mmap_min_addr)
1589 low_limit = mmap_min_addr;
1590 high_limit = info->high_limit;
1591retry:
1592 if (mas_empty_area(&mas, low_limit, high_limit - 1, length))
1593 return -ENOMEM;
1594
1595 gap = mas.index;
1596 gap += (info->align_offset - gap) & info->align_mask;
1597 tmp = mas_next(&mas, ULONG_MAX);
1598 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1599 if (vm_start_gap(tmp) < gap + length - 1) {
1600 low_limit = tmp->vm_end;
1601 mas_reset(&mas);
1602 goto retry;
1603 }
1604 } else {
1605 tmp = mas_prev(&mas, 0);
1606 if (tmp && vm_end_gap(tmp) > gap) {
1607 low_limit = vm_end_gap(tmp);
1608 mas_reset(&mas);
1609 goto retry;
1610 }
1611 }
1612
1613 return gap;
1614}
1615
1616/**
1617 * unmapped_area_topdown() - Find an area between the low_limit and the
1618 * high_limit with the correct alignment and offset at the highest available
1619 * address, all from @info. Note: current->mm is used for the search.
1620 *
1621 * @info: The unmapped area information including the range [low_limit -
1622 * high_limit), the alignment offset and mask.
1623 *
1624 * Return: A memory address or -ENOMEM.
1625 */
1626static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1627{
1628 unsigned long length, gap, gap_end;
1629 unsigned long low_limit, high_limit;
1630 struct vm_area_struct *tmp;
1631
1632 MA_STATE(mas, ¤t->mm->mm_mt, 0, 0);
1633 /* Adjust search length to account for worst case alignment overhead */
1634 length = info->length + info->align_mask;
1635 if (length < info->length)
1636 return -ENOMEM;
1637
1638 low_limit = info->low_limit;
1639 if (low_limit < mmap_min_addr)
1640 low_limit = mmap_min_addr;
1641 high_limit = info->high_limit;
1642retry:
1643 if (mas_empty_area_rev(&mas, low_limit, high_limit - 1, length))
1644 return -ENOMEM;
1645
1646 gap = mas.last + 1 - info->length;
1647 gap -= (gap - info->align_offset) & info->align_mask;
1648 gap_end = mas.last;
1649 tmp = mas_next(&mas, ULONG_MAX);
1650 if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1651 if (vm_start_gap(tmp) <= gap_end) {
1652 high_limit = vm_start_gap(tmp);
1653 mas_reset(&mas);
1654 goto retry;
1655 }
1656 } else {
1657 tmp = mas_prev(&mas, 0);
1658 if (tmp && vm_end_gap(tmp) > gap) {
1659 high_limit = tmp->vm_start;
1660 mas_reset(&mas);
1661 goto retry;
1662 }
1663 }
1664
1665 return gap;
1666}
1667
1668/*
1669 * Search for an unmapped address range.
1670 *
1671 * We are looking for a range that:
1672 * - does not intersect with any VMA;
1673 * - is contained within the [low_limit, high_limit) interval;
1674 * - is at least the desired size.
1675 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1676 */
1677unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1678{
1679 unsigned long addr;
1680
1681 if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1682 addr = unmapped_area_topdown(info);
1683 else
1684 addr = unmapped_area(info);
1685
1686 trace_vm_unmapped_area(addr, info);
1687 return addr;
1688}
1689
1690/* Get an address range which is currently unmapped.
1691 * For shmat() with addr=0.
1692 *
1693 * Ugly calling convention alert:
1694 * Return value with the low bits set means error value,
1695 * ie
1696 * if (ret & ~PAGE_MASK)
1697 * error = ret;
1698 *
1699 * This function "knows" that -ENOMEM has the bits set.
1700 */
1701unsigned long
1702generic_get_unmapped_area(struct file *filp, unsigned long addr,
1703 unsigned long len, unsigned long pgoff,
1704 unsigned long flags)
1705{
1706 struct mm_struct *mm = current->mm;
1707 struct vm_area_struct *vma, *prev;
1708 struct vm_unmapped_area_info info;
1709 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1710
1711 if (len > mmap_end - mmap_min_addr)
1712 return -ENOMEM;
1713
1714 if (flags & MAP_FIXED)
1715 return addr;
1716
1717 if (addr) {
1718 addr = PAGE_ALIGN(addr);
1719 vma = find_vma_prev(mm, addr, &prev);
1720 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1721 (!vma || addr + len <= vm_start_gap(vma)) &&
1722 (!prev || addr >= vm_end_gap(prev)))
1723 return addr;
1724 }
1725
1726 info.flags = 0;
1727 info.length = len;
1728 info.low_limit = mm->mmap_base;
1729 info.high_limit = mmap_end;
1730 info.align_mask = 0;
1731 info.align_offset = 0;
1732 return vm_unmapped_area(&info);
1733}
1734
1735#ifndef HAVE_ARCH_UNMAPPED_AREA
1736unsigned long
1737arch_get_unmapped_area(struct file *filp, unsigned long addr,
1738 unsigned long len, unsigned long pgoff,
1739 unsigned long flags)
1740{
1741 return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1742}
1743#endif
1744
1745/*
1746 * This mmap-allocator allocates new areas top-down from below the
1747 * stack's low limit (the base):
1748 */
1749unsigned long
1750generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1751 unsigned long len, unsigned long pgoff,
1752 unsigned long flags)
1753{
1754 struct vm_area_struct *vma, *prev;
1755 struct mm_struct *mm = current->mm;
1756 struct vm_unmapped_area_info info;
1757 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1758
1759 /* requested length too big for entire address space */
1760 if (len > mmap_end - mmap_min_addr)
1761 return -ENOMEM;
1762
1763 if (flags & MAP_FIXED)
1764 return addr;
1765
1766 /* requesting a specific address */
1767 if (addr) {
1768 addr = PAGE_ALIGN(addr);
1769 vma = find_vma_prev(mm, addr, &prev);
1770 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1771 (!vma || addr + len <= vm_start_gap(vma)) &&
1772 (!prev || addr >= vm_end_gap(prev)))
1773 return addr;
1774 }
1775
1776 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1777 info.length = len;
1778 info.low_limit = PAGE_SIZE;
1779 info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1780 info.align_mask = 0;
1781 info.align_offset = 0;
1782 addr = vm_unmapped_area(&info);
1783
1784 /*
1785 * A failed mmap() very likely causes application failure,
1786 * so fall back to the bottom-up function here. This scenario
1787 * can happen with large stack limits and large mmap()
1788 * allocations.
1789 */
1790 if (offset_in_page(addr)) {
1791 VM_BUG_ON(addr != -ENOMEM);
1792 info.flags = 0;
1793 info.low_limit = TASK_UNMAPPED_BASE;
1794 info.high_limit = mmap_end;
1795 addr = vm_unmapped_area(&info);
1796 }
1797
1798 return addr;
1799}
1800
1801#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1802unsigned long
1803arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1804 unsigned long len, unsigned long pgoff,
1805 unsigned long flags)
1806{
1807 return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1808}
1809#endif
1810
1811unsigned long
1812get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1813 unsigned long pgoff, unsigned long flags)
1814{
1815 unsigned long (*get_area)(struct file *, unsigned long,
1816 unsigned long, unsigned long, unsigned long);
1817
1818 unsigned long error = arch_mmap_check(addr, len, flags);
1819 if (error)
1820 return error;
1821
1822 /* Careful about overflows.. */
1823 if (len > TASK_SIZE)
1824 return -ENOMEM;
1825
1826 get_area = current->mm->get_unmapped_area;
1827 if (file) {
1828 if (file->f_op->get_unmapped_area)
1829 get_area = file->f_op->get_unmapped_area;
1830 } else if (flags & MAP_SHARED) {
1831 /*
1832 * mmap_region() will call shmem_zero_setup() to create a file,
1833 * so use shmem's get_unmapped_area in case it can be huge.
1834 */
1835 get_area = shmem_get_unmapped_area;
1836 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1837 /* Ensures that larger anonymous mappings are THP aligned. */
1838 get_area = thp_get_unmapped_area;
1839 }
1840
1841 /* Always treat pgoff as zero for anonymous memory. */
1842 if (!file)
1843 pgoff = 0;
1844
1845 addr = get_area(file, addr, len, pgoff, flags);
1846 if (IS_ERR_VALUE(addr))
1847 return addr;
1848
1849 if (addr > TASK_SIZE - len)
1850 return -ENOMEM;
1851 if (offset_in_page(addr))
1852 return -EINVAL;
1853
1854 error = security_mmap_addr(addr);
1855 return error ? error : addr;
1856}
1857
1858EXPORT_SYMBOL(get_unmapped_area);
1859
1860/**
1861 * find_vma_intersection() - Look up the first VMA which intersects the interval
1862 * @mm: The process address space.
1863 * @start_addr: The inclusive start user address.
1864 * @end_addr: The exclusive end user address.
1865 *
1866 * Returns: The first VMA within the provided range, %NULL otherwise. Assumes
1867 * start_addr < end_addr.
1868 */
1869struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1870 unsigned long start_addr,
1871 unsigned long end_addr)
1872{
1873 unsigned long index = start_addr;
1874
1875 mmap_assert_locked(mm);
1876 return mt_find(&mm->mm_mt, &index, end_addr - 1);
1877}
1878EXPORT_SYMBOL(find_vma_intersection);
1879
1880/**
1881 * find_vma() - Find the VMA for a given address, or the next VMA.
1882 * @mm: The mm_struct to check
1883 * @addr: The address
1884 *
1885 * Returns: The VMA associated with addr, or the next VMA.
1886 * May return %NULL in the case of no VMA at addr or above.
1887 */
1888struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1889{
1890 unsigned long index = addr;
1891
1892 mmap_assert_locked(mm);
1893 return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1894}
1895EXPORT_SYMBOL(find_vma);
1896
1897/**
1898 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1899 * set %pprev to the previous VMA, if any.
1900 * @mm: The mm_struct to check
1901 * @addr: The address
1902 * @pprev: The pointer to set to the previous VMA
1903 *
1904 * Note that RCU lock is missing here since the external mmap_lock() is used
1905 * instead.
1906 *
1907 * Returns: The VMA associated with @addr, or the next vma.
1908 * May return %NULL in the case of no vma at addr or above.
1909 */
1910struct vm_area_struct *
1911find_vma_prev(struct mm_struct *mm, unsigned long addr,
1912 struct vm_area_struct **pprev)
1913{
1914 struct vm_area_struct *vma;
1915 MA_STATE(mas, &mm->mm_mt, addr, addr);
1916
1917 vma = mas_walk(&mas);
1918 *pprev = mas_prev(&mas, 0);
1919 if (!vma)
1920 vma = mas_next(&mas, ULONG_MAX);
1921 return vma;
1922}
1923
1924/*
1925 * Verify that the stack growth is acceptable and
1926 * update accounting. This is shared with both the
1927 * grow-up and grow-down cases.
1928 */
1929static int acct_stack_growth(struct vm_area_struct *vma,
1930 unsigned long size, unsigned long grow)
1931{
1932 struct mm_struct *mm = vma->vm_mm;
1933 unsigned long new_start;
1934
1935 /* address space limit tests */
1936 if (!may_expand_vm(mm, vma->vm_flags, grow))
1937 return -ENOMEM;
1938
1939 /* Stack limit test */
1940 if (size > rlimit(RLIMIT_STACK))
1941 return -ENOMEM;
1942
1943 /* mlock limit tests */
1944 if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT))
1945 return -ENOMEM;
1946
1947 /* Check to ensure the stack will not grow into a hugetlb-only region */
1948 new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1949 vma->vm_end - size;
1950 if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1951 return -EFAULT;
1952
1953 /*
1954 * Overcommit.. This must be the final test, as it will
1955 * update security statistics.
1956 */
1957 if (security_vm_enough_memory_mm(mm, grow))
1958 return -ENOMEM;
1959
1960 return 0;
1961}
1962
1963#if defined(CONFIG_STACK_GROWSUP)
1964/*
1965 * PA-RISC uses this for its stack.
1966 * vma is the last one with address > vma->vm_end. Have to extend vma.
1967 */
1968static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1969{
1970 struct mm_struct *mm = vma->vm_mm;
1971 struct vm_area_struct *next;
1972 unsigned long gap_addr;
1973 int error = 0;
1974 MA_STATE(mas, &mm->mm_mt, vma->vm_start, address);
1975
1976 if (!(vma->vm_flags & VM_GROWSUP))
1977 return -EFAULT;
1978
1979 /* Guard against exceeding limits of the address space. */
1980 address &= PAGE_MASK;
1981 if (address >= (TASK_SIZE & PAGE_MASK))
1982 return -ENOMEM;
1983 address += PAGE_SIZE;
1984
1985 /* Enforce stack_guard_gap */
1986 gap_addr = address + stack_guard_gap;
1987
1988 /* Guard against overflow */
1989 if (gap_addr < address || gap_addr > TASK_SIZE)
1990 gap_addr = TASK_SIZE;
1991
1992 next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1993 if (next && vma_is_accessible(next)) {
1994 if (!(next->vm_flags & VM_GROWSUP))
1995 return -ENOMEM;
1996 /* Check that both stack segments have the same anon_vma? */
1997 }
1998
1999 if (next)
2000 mas_prev_range(&mas, address);
2001
2002 __mas_set_range(&mas, vma->vm_start, address - 1);
2003 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2004 return -ENOMEM;
2005
2006 /* We must make sure the anon_vma is allocated. */
2007 if (unlikely(anon_vma_prepare(vma))) {
2008 mas_destroy(&mas);
2009 return -ENOMEM;
2010 }
2011
2012 /* Lock the VMA before expanding to prevent concurrent page faults */
2013 vma_start_write(vma);
2014 /*
2015 * vma->vm_start/vm_end cannot change under us because the caller
2016 * is required to hold the mmap_lock in read mode. We need the
2017 * anon_vma lock to serialize against concurrent expand_stacks.
2018 */
2019 anon_vma_lock_write(vma->anon_vma);
2020
2021 /* Somebody else might have raced and expanded it already */
2022 if (address > vma->vm_end) {
2023 unsigned long size, grow;
2024
2025 size = address - vma->vm_start;
2026 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2027
2028 error = -ENOMEM;
2029 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2030 error = acct_stack_growth(vma, size, grow);
2031 if (!error) {
2032 /*
2033 * We only hold a shared mmap_lock lock here, so
2034 * we need to protect against concurrent vma
2035 * expansions. anon_vma_lock_write() doesn't
2036 * help here, as we don't guarantee that all
2037 * growable vmas in a mm share the same root
2038 * anon vma. So, we reuse mm->page_table_lock
2039 * to guard against concurrent vma expansions.
2040 */
2041 spin_lock(&mm->page_table_lock);
2042 if (vma->vm_flags & VM_LOCKED)
2043 mm->locked_vm += grow;
2044 vm_stat_account(mm, vma->vm_flags, grow);
2045 anon_vma_interval_tree_pre_update_vma(vma);
2046 vma->vm_end = address;
2047 /* Overwrite old entry in mtree. */
2048 mas_store_prealloc(&mas, vma);
2049 anon_vma_interval_tree_post_update_vma(vma);
2050 spin_unlock(&mm->page_table_lock);
2051
2052 perf_event_mmap(vma);
2053 }
2054 }
2055 }
2056 anon_vma_unlock_write(vma->anon_vma);
2057 mas_destroy(&mas);
2058 validate_mm(mm);
2059 return error;
2060}
2061#endif /* CONFIG_STACK_GROWSUP */
2062
2063/*
2064 * vma is the first one with address < vma->vm_start. Have to extend vma.
2065 * mmap_lock held for writing.
2066 */
2067int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2068{
2069 struct mm_struct *mm = vma->vm_mm;
2070 MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2071 struct vm_area_struct *prev;
2072 int error = 0;
2073
2074 if (!(vma->vm_flags & VM_GROWSDOWN))
2075 return -EFAULT;
2076
2077 address &= PAGE_MASK;
2078 if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
2079 return -EPERM;
2080
2081 /* Enforce stack_guard_gap */
2082 prev = mas_prev(&mas, 0);
2083 /* Check that both stack segments have the same anon_vma? */
2084 if (prev) {
2085 if (!(prev->vm_flags & VM_GROWSDOWN) &&
2086 vma_is_accessible(prev) &&
2087 (address - prev->vm_end < stack_guard_gap))
2088 return -ENOMEM;
2089 }
2090
2091 if (prev)
2092 mas_next_range(&mas, vma->vm_start);
2093
2094 __mas_set_range(&mas, address, vma->vm_end - 1);
2095 if (mas_preallocate(&mas, vma, GFP_KERNEL))
2096 return -ENOMEM;
2097
2098 /* We must make sure the anon_vma is allocated. */
2099 if (unlikely(anon_vma_prepare(vma))) {
2100 mas_destroy(&mas);
2101 return -ENOMEM;
2102 }
2103
2104 /* Lock the VMA before expanding to prevent concurrent page faults */
2105 vma_start_write(vma);
2106 /*
2107 * vma->vm_start/vm_end cannot change under us because the caller
2108 * is required to hold the mmap_lock in read mode. We need the
2109 * anon_vma lock to serialize against concurrent expand_stacks.
2110 */
2111 anon_vma_lock_write(vma->anon_vma);
2112
2113 /* Somebody else might have raced and expanded it already */
2114 if (address < vma->vm_start) {
2115 unsigned long size, grow;
2116
2117 size = vma->vm_end - address;
2118 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2119
2120 error = -ENOMEM;
2121 if (grow <= vma->vm_pgoff) {
2122 error = acct_stack_growth(vma, size, grow);
2123 if (!error) {
2124 /*
2125 * We only hold a shared mmap_lock lock here, so
2126 * we need to protect against concurrent vma
2127 * expansions. anon_vma_lock_write() doesn't
2128 * help here, as we don't guarantee that all
2129 * growable vmas in a mm share the same root
2130 * anon vma. So, we reuse mm->page_table_lock
2131 * to guard against concurrent vma expansions.
2132 */
2133 spin_lock(&mm->page_table_lock);
2134 if (vma->vm_flags & VM_LOCKED)
2135 mm->locked_vm += grow;
2136 vm_stat_account(mm, vma->vm_flags, grow);
2137 anon_vma_interval_tree_pre_update_vma(vma);
2138 vma->vm_start = address;
2139 vma->vm_pgoff -= grow;
2140 /* Overwrite old entry in mtree. */
2141 mas_store_prealloc(&mas, vma);
2142 anon_vma_interval_tree_post_update_vma(vma);
2143 spin_unlock(&mm->page_table_lock);
2144
2145 perf_event_mmap(vma);
2146 }
2147 }
2148 }
2149 anon_vma_unlock_write(vma->anon_vma);
2150 mas_destroy(&mas);
2151 validate_mm(mm);
2152 return error;
2153}
2154
2155/* enforced gap between the expanding stack and other mappings. */
2156unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2157
2158static int __init cmdline_parse_stack_guard_gap(char *p)
2159{
2160 unsigned long val;
2161 char *endptr;
2162
2163 val = simple_strtoul(p, &endptr, 10);
2164 if (!*endptr)
2165 stack_guard_gap = val << PAGE_SHIFT;
2166
2167 return 1;
2168}
2169__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2170
2171#ifdef CONFIG_STACK_GROWSUP
2172int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2173{
2174 return expand_upwards(vma, address);
2175}
2176
2177struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2178{
2179 struct vm_area_struct *vma, *prev;
2180
2181 addr &= PAGE_MASK;
2182 vma = find_vma_prev(mm, addr, &prev);
2183 if (vma && (vma->vm_start <= addr))
2184 return vma;
2185 if (!prev)
2186 return NULL;
2187 if (expand_stack_locked(prev, addr))
2188 return NULL;
2189 if (prev->vm_flags & VM_LOCKED)
2190 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2191 return prev;
2192}
2193#else
2194int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2195{
2196 return expand_downwards(vma, address);
2197}
2198
2199struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
2200{
2201 struct vm_area_struct *vma;
2202 unsigned long start;
2203
2204 addr &= PAGE_MASK;
2205 vma = find_vma(mm, addr);
2206 if (!vma)
2207 return NULL;
2208 if (vma->vm_start <= addr)
2209 return vma;
2210 start = vma->vm_start;
2211 if (expand_stack_locked(vma, addr))
2212 return NULL;
2213 if (vma->vm_flags & VM_LOCKED)
2214 populate_vma_page_range(vma, addr, start, NULL);
2215 return vma;
2216}
2217#endif
2218
2219#if defined(CONFIG_STACK_GROWSUP)
2220
2221#define vma_expand_up(vma,addr) expand_upwards(vma, addr)
2222#define vma_expand_down(vma, addr) (-EFAULT)
2223
2224#else
2225
2226#define vma_expand_up(vma,addr) (-EFAULT)
2227#define vma_expand_down(vma, addr) expand_downwards(vma, addr)
2228
2229#endif
2230
2231/*
2232 * expand_stack(): legacy interface for page faulting. Don't use unless
2233 * you have to.
2234 *
2235 * This is called with the mm locked for reading, drops the lock, takes
2236 * the lock for writing, tries to look up a vma again, expands it if
2237 * necessary, and downgrades the lock to reading again.
2238 *
2239 * If no vma is found or it can't be expanded, it returns NULL and has
2240 * dropped the lock.
2241 */
2242struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
2243{
2244 struct vm_area_struct *vma, *prev;
2245
2246 mmap_read_unlock(mm);
2247 if (mmap_write_lock_killable(mm))
2248 return NULL;
2249
2250 vma = find_vma_prev(mm, addr, &prev);
2251 if (vma && vma->vm_start <= addr)
2252 goto success;
2253
2254 if (prev && !vma_expand_up(prev, addr)) {
2255 vma = prev;
2256 goto success;
2257 }
2258
2259 if (vma && !vma_expand_down(vma, addr))
2260 goto success;
2261
2262 mmap_write_unlock(mm);
2263 return NULL;
2264
2265success:
2266 mmap_write_downgrade(mm);
2267 return vma;
2268}
2269
2270/*
2271 * Ok - we have the memory areas we should free on a maple tree so release them,
2272 * and do the vma updates.
2273 *
2274 * Called with the mm semaphore held.
2275 */
2276static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2277{
2278 unsigned long nr_accounted = 0;
2279 struct vm_area_struct *vma;
2280
2281 /* Update high watermark before we lower total_vm */
2282 update_hiwater_vm(mm);
2283 mas_for_each(mas, vma, ULONG_MAX) {
2284 long nrpages = vma_pages(vma);
2285
2286 if (vma->vm_flags & VM_ACCOUNT)
2287 nr_accounted += nrpages;
2288 vm_stat_account(mm, vma->vm_flags, -nrpages);
2289 remove_vma(vma, false);
2290 }
2291 vm_unacct_memory(nr_accounted);
2292}
2293
2294/*
2295 * Get rid of page table information in the indicated region.
2296 *
2297 * Called with the mm semaphore held.
2298 */
2299static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
2300 struct vm_area_struct *vma, struct vm_area_struct *prev,
2301 struct vm_area_struct *next, unsigned long start,
2302 unsigned long end, unsigned long tree_end, bool mm_wr_locked)
2303{
2304 struct mmu_gather tlb;
2305 unsigned long mt_start = mas->index;
2306
2307 lru_add_drain();
2308 tlb_gather_mmu(&tlb, mm);
2309 update_hiwater_rss(mm);
2310 unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
2311 mas_set(mas, mt_start);
2312 free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2313 next ? next->vm_start : USER_PGTABLES_CEILING,
2314 mm_wr_locked);
2315 tlb_finish_mmu(&tlb);
2316}
2317
2318/*
2319 * __split_vma() bypasses sysctl_max_map_count checking. We use this where it
2320 * has already been checked or doesn't make sense to fail.
2321 * VMA Iterator will point to the end VMA.
2322 */
2323static int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2324 unsigned long addr, int new_below)
2325{
2326 struct vma_prepare vp;
2327 struct vm_area_struct *new;
2328 int err;
2329
2330 WARN_ON(vma->vm_start >= addr);
2331 WARN_ON(vma->vm_end <= addr);
2332
2333 if (vma->vm_ops && vma->vm_ops->may_split) {
2334 err = vma->vm_ops->may_split(vma, addr);
2335 if (err)
2336 return err;
2337 }
2338
2339 new = vm_area_dup(vma);
2340 if (!new)
2341 return -ENOMEM;
2342
2343 if (new_below) {
2344 new->vm_end = addr;
2345 } else {
2346 new->vm_start = addr;
2347 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2348 }
2349
2350 err = -ENOMEM;
2351 vma_iter_config(vmi, new->vm_start, new->vm_end);
2352 if (vma_iter_prealloc(vmi, new))
2353 goto out_free_vma;
2354
2355 err = vma_dup_policy(vma, new);
2356 if (err)
2357 goto out_free_vmi;
2358
2359 err = anon_vma_clone(new, vma);
2360 if (err)
2361 goto out_free_mpol;
2362
2363 if (new->vm_file)
2364 get_file(new->vm_file);
2365
2366 if (new->vm_ops && new->vm_ops->open)
2367 new->vm_ops->open(new);
2368
2369 vma_start_write(vma);
2370 vma_start_write(new);
2371
2372 init_vma_prep(&vp, vma);
2373 vp.insert = new;
2374 vma_prepare(&vp);
2375 vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
2376
2377 if (new_below) {
2378 vma->vm_start = addr;
2379 vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
2380 } else {
2381 vma->vm_end = addr;
2382 }
2383
2384 /* vma_complete stores the new vma */
2385 vma_complete(&vp, vmi, vma->vm_mm);
2386
2387 /* Success. */
2388 if (new_below)
2389 vma_next(vmi);
2390 return 0;
2391
2392out_free_mpol:
2393 mpol_put(vma_policy(new));
2394out_free_vmi:
2395 vma_iter_free(vmi);
2396out_free_vma:
2397 vm_area_free(new);
2398 return err;
2399}
2400
2401/*
2402 * Split a vma into two pieces at address 'addr', a new vma is allocated
2403 * either for the first part or the tail.
2404 */
2405static int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2406 unsigned long addr, int new_below)
2407{
2408 if (vma->vm_mm->map_count >= sysctl_max_map_count)
2409 return -ENOMEM;
2410
2411 return __split_vma(vmi, vma, addr, new_below);
2412}
2413
2414/*
2415 * We are about to modify one or multiple of a VMA's flags, policy, userfaultfd
2416 * context and anonymous VMA name within the range [start, end).
2417 *
2418 * As a result, we might be able to merge the newly modified VMA range with an
2419 * adjacent VMA with identical properties.
2420 *
2421 * If no merge is possible and the range does not span the entirety of the VMA,
2422 * we then need to split the VMA to accommodate the change.
2423 *
2424 * The function returns either the merged VMA, the original VMA if a split was
2425 * required instead, or an error if the split failed.
2426 */
2427struct vm_area_struct *vma_modify(struct vma_iterator *vmi,
2428 struct vm_area_struct *prev,
2429 struct vm_area_struct *vma,
2430 unsigned long start, unsigned long end,
2431 unsigned long vm_flags,
2432 struct mempolicy *policy,
2433 struct vm_userfaultfd_ctx uffd_ctx,
2434 struct anon_vma_name *anon_name)
2435{
2436 pgoff_t pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
2437 struct vm_area_struct *merged;
2438
2439 merged = vma_merge(vmi, prev, vma, start, end, vm_flags,
2440 pgoff, policy, uffd_ctx, anon_name);
2441 if (merged)
2442 return merged;
2443
2444 if (vma->vm_start < start) {
2445 int err = split_vma(vmi, vma, start, 1);
2446
2447 if (err)
2448 return ERR_PTR(err);
2449 }
2450
2451 if (vma->vm_end > end) {
2452 int err = split_vma(vmi, vma, end, 0);
2453
2454 if (err)
2455 return ERR_PTR(err);
2456 }
2457
2458 return vma;
2459}
2460
2461/*
2462 * Attempt to merge a newly mapped VMA with those adjacent to it. The caller
2463 * must ensure that [start, end) does not overlap any existing VMA.
2464 */
2465static struct vm_area_struct
2466*vma_merge_new_vma(struct vma_iterator *vmi, struct vm_area_struct *prev,
2467 struct vm_area_struct *vma, unsigned long start,
2468 unsigned long end, pgoff_t pgoff)
2469{
2470 return vma_merge(vmi, prev, vma, start, end, vma->vm_flags, pgoff,
2471 vma_policy(vma), vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2472}
2473
2474/*
2475 * Expand vma by delta bytes, potentially merging with an immediately adjacent
2476 * VMA with identical properties.
2477 */
2478struct vm_area_struct *vma_merge_extend(struct vma_iterator *vmi,
2479 struct vm_area_struct *vma,
2480 unsigned long delta)
2481{
2482 pgoff_t pgoff = vma->vm_pgoff + vma_pages(vma);
2483
2484 /* vma is specified as prev, so case 1 or 2 will apply. */
2485 return vma_merge(vmi, vma, vma, vma->vm_end, vma->vm_end + delta,
2486 vma->vm_flags, pgoff, vma_policy(vma),
2487 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2488}
2489
2490/*
2491 * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
2492 * @vmi: The vma iterator
2493 * @vma: The starting vm_area_struct
2494 * @mm: The mm_struct
2495 * @start: The aligned start address to munmap.
2496 * @end: The aligned end address to munmap.
2497 * @uf: The userfaultfd list_head
2498 * @unlock: Set to true to drop the mmap_lock. unlocking only happens on
2499 * success.
2500 *
2501 * Return: 0 on success and drops the lock if so directed, error and leaves the
2502 * lock held otherwise.
2503 */
2504static int
2505do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2506 struct mm_struct *mm, unsigned long start,
2507 unsigned long end, struct list_head *uf, bool unlock)
2508{
2509 struct vm_area_struct *prev, *next = NULL;
2510 struct maple_tree mt_detach;
2511 int count = 0;
2512 int error = -ENOMEM;
2513 unsigned long locked_vm = 0;
2514 MA_STATE(mas_detach, &mt_detach, 0, 0);
2515 mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
2516 mt_on_stack(mt_detach);
2517
2518 /*
2519 * If we need to split any vma, do it now to save pain later.
2520 *
2521 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2522 * unmapped vm_area_struct will remain in use: so lower split_vma
2523 * places tmp vma above, and higher split_vma places tmp vma below.
2524 */
2525
2526 /* Does it split the first one? */
2527 if (start > vma->vm_start) {
2528
2529 /*
2530 * Make sure that map_count on return from munmap() will
2531 * not exceed its limit; but let map_count go just above
2532 * its limit temporarily, to help free resources as expected.
2533 */
2534 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2535 goto map_count_exceeded;
2536
2537 error = __split_vma(vmi, vma, start, 1);
2538 if (error)
2539 goto start_split_failed;
2540 }
2541
2542 /*
2543 * Detach a range of VMAs from the mm. Using next as a temp variable as
2544 * it is always overwritten.
2545 */
2546 next = vma;
2547 do {
2548 /* Does it split the end? */
2549 if (next->vm_end > end) {
2550 error = __split_vma(vmi, next, end, 0);
2551 if (error)
2552 goto end_split_failed;
2553 }
2554 vma_start_write(next);
2555 mas_set(&mas_detach, count);
2556 error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
2557 if (error)
2558 goto munmap_gather_failed;
2559 vma_mark_detached(next, true);
2560 if (next->vm_flags & VM_LOCKED)
2561 locked_vm += vma_pages(next);
2562
2563 count++;
2564 if (unlikely(uf)) {
2565 /*
2566 * If userfaultfd_unmap_prep returns an error the vmas
2567 * will remain split, but userland will get a
2568 * highly unexpected error anyway. This is no
2569 * different than the case where the first of the two
2570 * __split_vma fails, but we don't undo the first
2571 * split, despite we could. This is unlikely enough
2572 * failure that it's not worth optimizing it for.
2573 */
2574 error = userfaultfd_unmap_prep(next, start, end, uf);
2575
2576 if (error)
2577 goto userfaultfd_error;
2578 }
2579#ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2580 BUG_ON(next->vm_start < start);
2581 BUG_ON(next->vm_start > end);
2582#endif
2583 } for_each_vma_range(*vmi, next, end);
2584
2585#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2586 /* Make sure no VMAs are about to be lost. */
2587 {
2588 MA_STATE(test, &mt_detach, 0, 0);
2589 struct vm_area_struct *vma_mas, *vma_test;
2590 int test_count = 0;
2591
2592 vma_iter_set(vmi, start);
2593 rcu_read_lock();
2594 vma_test = mas_find(&test, count - 1);
2595 for_each_vma_range(*vmi, vma_mas, end) {
2596 BUG_ON(vma_mas != vma_test);
2597 test_count++;
2598 vma_test = mas_next(&test, count - 1);
2599 }
2600 rcu_read_unlock();
2601 BUG_ON(count != test_count);
2602 }
2603#endif
2604
2605 while (vma_iter_addr(vmi) > start)
2606 vma_iter_prev_range(vmi);
2607
2608 error = vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL);
2609 if (error)
2610 goto clear_tree_failed;
2611
2612 /* Point of no return */
2613 mm->locked_vm -= locked_vm;
2614 mm->map_count -= count;
2615 if (unlock)
2616 mmap_write_downgrade(mm);
2617
2618 prev = vma_iter_prev_range(vmi);
2619 next = vma_next(vmi);
2620 if (next)
2621 vma_iter_prev_range(vmi);
2622
2623 /*
2624 * We can free page tables without write-locking mmap_lock because VMAs
2625 * were isolated before we downgraded mmap_lock.
2626 */
2627 mas_set(&mas_detach, 1);
2628 unmap_region(mm, &mas_detach, vma, prev, next, start, end, count,
2629 !unlock);
2630 /* Statistics and freeing VMAs */
2631 mas_set(&mas_detach, 0);
2632 remove_mt(mm, &mas_detach);
2633 validate_mm(mm);
2634 if (unlock)
2635 mmap_read_unlock(mm);
2636
2637 __mt_destroy(&mt_detach);
2638 return 0;
2639
2640clear_tree_failed:
2641userfaultfd_error:
2642munmap_gather_failed:
2643end_split_failed:
2644 mas_set(&mas_detach, 0);
2645 mas_for_each(&mas_detach, next, end)
2646 vma_mark_detached(next, false);
2647
2648 __mt_destroy(&mt_detach);
2649start_split_failed:
2650map_count_exceeded:
2651 validate_mm(mm);
2652 return error;
2653}
2654
2655/*
2656 * do_vmi_munmap() - munmap a given range.
2657 * @vmi: The vma iterator
2658 * @mm: The mm_struct
2659 * @start: The start address to munmap
2660 * @len: The length of the range to munmap
2661 * @uf: The userfaultfd list_head
2662 * @unlock: set to true if the user wants to drop the mmap_lock on success
2663 *
2664 * This function takes a @mas that is either pointing to the previous VMA or set
2665 * to MA_START and sets it up to remove the mapping(s). The @len will be
2666 * aligned and any arch_unmap work will be preformed.
2667 *
2668 * Return: 0 on success and drops the lock if so directed, error and leaves the
2669 * lock held otherwise.
2670 */
2671int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
2672 unsigned long start, size_t len, struct list_head *uf,
2673 bool unlock)
2674{
2675 unsigned long end;
2676 struct vm_area_struct *vma;
2677
2678 if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2679 return -EINVAL;
2680
2681 end = start + PAGE_ALIGN(len);
2682 if (end == start)
2683 return -EINVAL;
2684
2685 /* arch_unmap() might do unmaps itself. */
2686 arch_unmap(mm, start, end);
2687
2688 /* Find the first overlapping VMA */
2689 vma = vma_find(vmi, end);
2690 if (!vma) {
2691 if (unlock)
2692 mmap_write_unlock(mm);
2693 return 0;
2694 }
2695
2696 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
2697}
2698
2699/* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2700 * @mm: The mm_struct
2701 * @start: The start address to munmap
2702 * @len: The length to be munmapped.
2703 * @uf: The userfaultfd list_head
2704 *
2705 * Return: 0 on success, error otherwise.
2706 */
2707int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2708 struct list_head *uf)
2709{
2710 VMA_ITERATOR(vmi, mm, start);
2711
2712 return do_vmi_munmap(&vmi, mm, start, len, uf, false);
2713}
2714
2715unsigned long mmap_region(struct file *file, unsigned long addr,
2716 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2717 struct list_head *uf)
2718{
2719 struct mm_struct *mm = current->mm;
2720 struct vm_area_struct *vma = NULL;
2721 struct vm_area_struct *next, *prev, *merge;
2722 pgoff_t pglen = len >> PAGE_SHIFT;
2723 unsigned long charged = 0;
2724 unsigned long end = addr + len;
2725 unsigned long merge_start = addr, merge_end = end;
2726 bool writable_file_mapping = false;
2727 pgoff_t vm_pgoff;
2728 int error;
2729 VMA_ITERATOR(vmi, mm, addr);
2730
2731 /* Check against address space limit. */
2732 if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2733 unsigned long nr_pages;
2734
2735 /*
2736 * MAP_FIXED may remove pages of mappings that intersects with
2737 * requested mapping. Account for the pages it would unmap.
2738 */
2739 nr_pages = count_vma_pages_range(mm, addr, end);
2740
2741 if (!may_expand_vm(mm, vm_flags,
2742 (len >> PAGE_SHIFT) - nr_pages))
2743 return -ENOMEM;
2744 }
2745
2746 /* Unmap any existing mapping in the area */
2747 if (do_vmi_munmap(&vmi, mm, addr, len, uf, false))
2748 return -ENOMEM;
2749
2750 /*
2751 * Private writable mapping: check memory availability
2752 */
2753 if (accountable_mapping(file, vm_flags)) {
2754 charged = len >> PAGE_SHIFT;
2755 if (security_vm_enough_memory_mm(mm, charged))
2756 return -ENOMEM;
2757 vm_flags |= VM_ACCOUNT;
2758 }
2759
2760 next = vma_next(&vmi);
2761 prev = vma_prev(&vmi);
2762 if (vm_flags & VM_SPECIAL) {
2763 if (prev)
2764 vma_iter_next_range(&vmi);
2765 goto cannot_expand;
2766 }
2767
2768 /* Attempt to expand an old mapping */
2769 /* Check next */
2770 if (next && next->vm_start == end && !vma_policy(next) &&
2771 can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2772 NULL_VM_UFFD_CTX, NULL)) {
2773 merge_end = next->vm_end;
2774 vma = next;
2775 vm_pgoff = next->vm_pgoff - pglen;
2776 }
2777
2778 /* Check prev */
2779 if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2780 (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2781 pgoff, vma->vm_userfaultfd_ctx, NULL) :
2782 can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2783 NULL_VM_UFFD_CTX, NULL))) {
2784 merge_start = prev->vm_start;
2785 vma = prev;
2786 vm_pgoff = prev->vm_pgoff;
2787 } else if (prev) {
2788 vma_iter_next_range(&vmi);
2789 }
2790
2791 /* Actually expand, if possible */
2792 if (vma &&
2793 !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) {
2794 khugepaged_enter_vma(vma, vm_flags);
2795 goto expanded;
2796 }
2797
2798 if (vma == prev)
2799 vma_iter_set(&vmi, addr);
2800cannot_expand:
2801
2802 /*
2803 * Determine the object being mapped and call the appropriate
2804 * specific mapper. the address has already been validated, but
2805 * not unmapped, but the maps are removed from the list.
2806 */
2807 vma = vm_area_alloc(mm);
2808 if (!vma) {
2809 error = -ENOMEM;
2810 goto unacct_error;
2811 }
2812
2813 vma_iter_config(&vmi, addr, end);
2814 vma_set_range(vma, addr, end, pgoff);
2815 vm_flags_init(vma, vm_flags);
2816 vma->vm_page_prot = vm_get_page_prot(vm_flags);
2817
2818 if (file) {
2819 vma->vm_file = get_file(file);
2820 error = call_mmap(file, vma);
2821 if (error)
2822 goto unmap_and_free_vma;
2823
2824 if (vma_is_shared_maywrite(vma)) {
2825 error = mapping_map_writable(file->f_mapping);
2826 if (error)
2827 goto close_and_free_vma;
2828
2829 writable_file_mapping = true;
2830 }
2831
2832 /*
2833 * Expansion is handled above, merging is handled below.
2834 * Drivers should not alter the address of the VMA.
2835 */
2836 error = -EINVAL;
2837 if (WARN_ON((addr != vma->vm_start)))
2838 goto close_and_free_vma;
2839
2840 vma_iter_config(&vmi, addr, end);
2841 /*
2842 * If vm_flags changed after call_mmap(), we should try merge
2843 * vma again as we may succeed this time.
2844 */
2845 if (unlikely(vm_flags != vma->vm_flags && prev)) {
2846 merge = vma_merge_new_vma(&vmi, prev, vma,
2847 vma->vm_start, vma->vm_end,
2848 vma->vm_pgoff);
2849 if (merge) {
2850 /*
2851 * ->mmap() can change vma->vm_file and fput
2852 * the original file. So fput the vma->vm_file
2853 * here or we would add an extra fput for file
2854 * and cause general protection fault
2855 * ultimately.
2856 */
2857 fput(vma->vm_file);
2858 vm_area_free(vma);
2859 vma = merge;
2860 /* Update vm_flags to pick up the change. */
2861 vm_flags = vma->vm_flags;
2862 goto unmap_writable;
2863 }
2864 }
2865
2866 vm_flags = vma->vm_flags;
2867 } else if (vm_flags & VM_SHARED) {
2868 error = shmem_zero_setup(vma);
2869 if (error)
2870 goto free_vma;
2871 } else {
2872 vma_set_anonymous(vma);
2873 }
2874
2875 if (map_deny_write_exec(vma, vma->vm_flags)) {
2876 error = -EACCES;
2877 goto close_and_free_vma;
2878 }
2879
2880 /* Allow architectures to sanity-check the vm_flags */
2881 error = -EINVAL;
2882 if (!arch_validate_flags(vma->vm_flags))
2883 goto close_and_free_vma;
2884
2885 error = -ENOMEM;
2886 if (vma_iter_prealloc(&vmi, vma))
2887 goto close_and_free_vma;
2888
2889 /* Lock the VMA since it is modified after insertion into VMA tree */
2890 vma_start_write(vma);
2891 vma_iter_store(&vmi, vma);
2892 mm->map_count++;
2893 vma_link_file(vma);
2894
2895 /*
2896 * vma_merge() calls khugepaged_enter_vma() either, the below
2897 * call covers the non-merge case.
2898 */
2899 khugepaged_enter_vma(vma, vma->vm_flags);
2900
2901 /* Once vma denies write, undo our temporary denial count */
2902unmap_writable:
2903 if (writable_file_mapping)
2904 mapping_unmap_writable(file->f_mapping);
2905 file = vma->vm_file;
2906 ksm_add_vma(vma);
2907expanded:
2908 perf_event_mmap(vma);
2909
2910 vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2911 if (vm_flags & VM_LOCKED) {
2912 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2913 is_vm_hugetlb_page(vma) ||
2914 vma == get_gate_vma(current->mm))
2915 vm_flags_clear(vma, VM_LOCKED_MASK);
2916 else
2917 mm->locked_vm += (len >> PAGE_SHIFT);
2918 }
2919
2920 if (file)
2921 uprobe_mmap(vma);
2922
2923 /*
2924 * New (or expanded) vma always get soft dirty status.
2925 * Otherwise user-space soft-dirty page tracker won't
2926 * be able to distinguish situation when vma area unmapped,
2927 * then new mapped in-place (which must be aimed as
2928 * a completely new data area).
2929 */
2930 vm_flags_set(vma, VM_SOFTDIRTY);
2931
2932 vma_set_page_prot(vma);
2933
2934 validate_mm(mm);
2935 return addr;
2936
2937close_and_free_vma:
2938 if (file && vma->vm_ops && vma->vm_ops->close)
2939 vma->vm_ops->close(vma);
2940
2941 if (file || vma->vm_file) {
2942unmap_and_free_vma:
2943 fput(vma->vm_file);
2944 vma->vm_file = NULL;
2945
2946 vma_iter_set(&vmi, vma->vm_end);
2947 /* Undo any partial mapping done by a device driver. */
2948 unmap_region(mm, &vmi.mas, vma, prev, next, vma->vm_start,
2949 vma->vm_end, vma->vm_end, true);
2950 }
2951 if (writable_file_mapping)
2952 mapping_unmap_writable(file->f_mapping);
2953free_vma:
2954 vm_area_free(vma);
2955unacct_error:
2956 if (charged)
2957 vm_unacct_memory(charged);
2958 validate_mm(mm);
2959 return error;
2960}
2961
2962static int __vm_munmap(unsigned long start, size_t len, bool unlock)
2963{
2964 int ret;
2965 struct mm_struct *mm = current->mm;
2966 LIST_HEAD(uf);
2967 VMA_ITERATOR(vmi, mm, start);
2968
2969 if (mmap_write_lock_killable(mm))
2970 return -EINTR;
2971
2972 ret = do_vmi_munmap(&vmi, mm, start, len, &uf, unlock);
2973 if (ret || !unlock)
2974 mmap_write_unlock(mm);
2975
2976 userfaultfd_unmap_complete(mm, &uf);
2977 return ret;
2978}
2979
2980int vm_munmap(unsigned long start, size_t len)
2981{
2982 return __vm_munmap(start, len, false);
2983}
2984EXPORT_SYMBOL(vm_munmap);
2985
2986SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2987{
2988 addr = untagged_addr(addr);
2989 return __vm_munmap(addr, len, true);
2990}
2991
2992
2993/*
2994 * Emulation of deprecated remap_file_pages() syscall.
2995 */
2996SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2997 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2998{
2999
3000 struct mm_struct *mm = current->mm;
3001 struct vm_area_struct *vma;
3002 unsigned long populate = 0;
3003 unsigned long ret = -EINVAL;
3004 struct file *file;
3005
3006 pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
3007 current->comm, current->pid);
3008
3009 if (prot)
3010 return ret;
3011 start = start & PAGE_MASK;
3012 size = size & PAGE_MASK;
3013
3014 if (start + size <= start)
3015 return ret;
3016
3017 /* Does pgoff wrap? */
3018 if (pgoff + (size >> PAGE_SHIFT) < pgoff)
3019 return ret;
3020
3021 if (mmap_write_lock_killable(mm))
3022 return -EINTR;
3023
3024 vma = vma_lookup(mm, start);
3025
3026 if (!vma || !(vma->vm_flags & VM_SHARED))
3027 goto out;
3028
3029 if (start + size > vma->vm_end) {
3030 VMA_ITERATOR(vmi, mm, vma->vm_end);
3031 struct vm_area_struct *next, *prev = vma;
3032
3033 for_each_vma_range(vmi, next, start + size) {
3034 /* hole between vmas ? */
3035 if (next->vm_start != prev->vm_end)
3036 goto out;
3037
3038 if (next->vm_file != vma->vm_file)
3039 goto out;
3040
3041 if (next->vm_flags != vma->vm_flags)
3042 goto out;
3043
3044 if (start + size <= next->vm_end)
3045 break;
3046
3047 prev = next;
3048 }
3049
3050 if (!next)
3051 goto out;
3052 }
3053
3054 prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3055 prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3056 prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3057
3058 flags &= MAP_NONBLOCK;
3059 flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3060 if (vma->vm_flags & VM_LOCKED)
3061 flags |= MAP_LOCKED;
3062
3063 file = get_file(vma->vm_file);
3064 ret = do_mmap(vma->vm_file, start, size,
3065 prot, flags, 0, pgoff, &populate, NULL);
3066 fput(file);
3067out:
3068 mmap_write_unlock(mm);
3069 if (populate)
3070 mm_populate(ret, populate);
3071 if (!IS_ERR_VALUE(ret))
3072 ret = 0;
3073 return ret;
3074}
3075
3076/*
3077 * do_vma_munmap() - Unmap a full or partial vma.
3078 * @vmi: The vma iterator pointing at the vma
3079 * @vma: The first vma to be munmapped
3080 * @start: the start of the address to unmap
3081 * @end: The end of the address to unmap
3082 * @uf: The userfaultfd list_head
3083 * @unlock: Drop the lock on success
3084 *
3085 * unmaps a VMA mapping when the vma iterator is already in position.
3086 * Does not handle alignment.
3087 *
3088 * Return: 0 on success drops the lock of so directed, error on failure and will
3089 * still hold the lock.
3090 */
3091int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
3092 unsigned long start, unsigned long end, struct list_head *uf,
3093 bool unlock)
3094{
3095 struct mm_struct *mm = vma->vm_mm;
3096
3097 arch_unmap(mm, start, end);
3098 return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
3099}
3100
3101/*
3102 * do_brk_flags() - Increase the brk vma if the flags match.
3103 * @vmi: The vma iterator
3104 * @addr: The start address
3105 * @len: The length of the increase
3106 * @vma: The vma,
3107 * @flags: The VMA Flags
3108 *
3109 * Extend the brk VMA from addr to addr + len. If the VMA is NULL or the flags
3110 * do not match then create a new anonymous VMA. Eventually we may be able to
3111 * do some brk-specific accounting here.
3112 */
3113static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
3114 unsigned long addr, unsigned long len, unsigned long flags)
3115{
3116 struct mm_struct *mm = current->mm;
3117 struct vma_prepare vp;
3118
3119 /*
3120 * Check against address space limits by the changed size
3121 * Note: This happens *after* clearing old mappings in some code paths.
3122 */
3123 flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3124 if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3125 return -ENOMEM;
3126
3127 if (mm->map_count > sysctl_max_map_count)
3128 return -ENOMEM;
3129
3130 if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3131 return -ENOMEM;
3132
3133 /*
3134 * Expand the existing vma if possible; Note that singular lists do not
3135 * occur after forking, so the expand will only happen on new VMAs.
3136 */
3137 if (vma && vma->vm_end == addr && !vma_policy(vma) &&
3138 can_vma_merge_after(vma, flags, NULL, NULL,
3139 addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
3140 vma_iter_config(vmi, vma->vm_start, addr + len);
3141 if (vma_iter_prealloc(vmi, vma))
3142 goto unacct_fail;
3143
3144 vma_start_write(vma);
3145
3146 init_vma_prep(&vp, vma);
3147 vma_prepare(&vp);
3148 vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
3149 vma->vm_end = addr + len;
3150 vm_flags_set(vma, VM_SOFTDIRTY);
3151 vma_iter_store(vmi, vma);
3152
3153 vma_complete(&vp, vmi, mm);
3154 khugepaged_enter_vma(vma, flags);
3155 goto out;
3156 }
3157
3158 if (vma)
3159 vma_iter_next_range(vmi);
3160 /* create a vma struct for an anonymous mapping */
3161 vma = vm_area_alloc(mm);
3162 if (!vma)
3163 goto unacct_fail;
3164
3165 vma_set_anonymous(vma);
3166 vma_set_range(vma, addr, addr + len, addr >> PAGE_SHIFT);
3167 vm_flags_init(vma, flags);
3168 vma->vm_page_prot = vm_get_page_prot(flags);
3169 vma_start_write(vma);
3170 if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
3171 goto mas_store_fail;
3172
3173 mm->map_count++;
3174 validate_mm(mm);
3175 ksm_add_vma(vma);
3176out:
3177 perf_event_mmap(vma);
3178 mm->total_vm += len >> PAGE_SHIFT;
3179 mm->data_vm += len >> PAGE_SHIFT;
3180 if (flags & VM_LOCKED)
3181 mm->locked_vm += (len >> PAGE_SHIFT);
3182 vm_flags_set(vma, VM_SOFTDIRTY);
3183 return 0;
3184
3185mas_store_fail:
3186 vm_area_free(vma);
3187unacct_fail:
3188 vm_unacct_memory(len >> PAGE_SHIFT);
3189 return -ENOMEM;
3190}
3191
3192int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3193{
3194 struct mm_struct *mm = current->mm;
3195 struct vm_area_struct *vma = NULL;
3196 unsigned long len;
3197 int ret;
3198 bool populate;
3199 LIST_HEAD(uf);
3200 VMA_ITERATOR(vmi, mm, addr);
3201
3202 len = PAGE_ALIGN(request);
3203 if (len < request)
3204 return -ENOMEM;
3205 if (!len)
3206 return 0;
3207
3208 /* Until we need other flags, refuse anything except VM_EXEC. */
3209 if ((flags & (~VM_EXEC)) != 0)
3210 return -EINVAL;
3211
3212 if (mmap_write_lock_killable(mm))
3213 return -EINTR;
3214
3215 ret = check_brk_limits(addr, len);
3216 if (ret)
3217 goto limits_failed;
3218
3219 ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
3220 if (ret)
3221 goto munmap_failed;
3222
3223 vma = vma_prev(&vmi);
3224 ret = do_brk_flags(&vmi, vma, addr, len, flags);
3225 populate = ((mm->def_flags & VM_LOCKED) != 0);
3226 mmap_write_unlock(mm);
3227 userfaultfd_unmap_complete(mm, &uf);
3228 if (populate && !ret)
3229 mm_populate(addr, len);
3230 return ret;
3231
3232munmap_failed:
3233limits_failed:
3234 mmap_write_unlock(mm);
3235 return ret;
3236}
3237EXPORT_SYMBOL(vm_brk_flags);
3238
3239/* Release all mmaps. */
3240void exit_mmap(struct mm_struct *mm)
3241{
3242 struct mmu_gather tlb;
3243 struct vm_area_struct *vma;
3244 unsigned long nr_accounted = 0;
3245 MA_STATE(mas, &mm->mm_mt, 0, 0);
3246 int count = 0;
3247
3248 /* mm's last user has gone, and its about to be pulled down */
3249 mmu_notifier_release(mm);
3250
3251 mmap_read_lock(mm);
3252 arch_exit_mmap(mm);
3253
3254 vma = mas_find(&mas, ULONG_MAX);
3255 if (!vma || unlikely(xa_is_zero(vma))) {
3256 /* Can happen if dup_mmap() received an OOM */
3257 mmap_read_unlock(mm);
3258 mmap_write_lock(mm);
3259 goto destroy;
3260 }
3261
3262 lru_add_drain();
3263 flush_cache_mm(mm);
3264 tlb_gather_mmu_fullmm(&tlb, mm);
3265 /* update_hiwater_rss(mm) here? but nobody should be looking */
3266 /* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3267 unmap_vmas(&tlb, &mas, vma, 0, ULONG_MAX, ULONG_MAX, false);
3268 mmap_read_unlock(mm);
3269
3270 /*
3271 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3272 * because the memory has been already freed.
3273 */
3274 set_bit(MMF_OOM_SKIP, &mm->flags);
3275 mmap_write_lock(mm);
3276 mt_clear_in_rcu(&mm->mm_mt);
3277 mas_set(&mas, vma->vm_end);
3278 free_pgtables(&tlb, &mas, vma, FIRST_USER_ADDRESS,
3279 USER_PGTABLES_CEILING, true);
3280 tlb_finish_mmu(&tlb);
3281
3282 /*
3283 * Walk the list again, actually closing and freeing it, with preemption
3284 * enabled, without holding any MM locks besides the unreachable
3285 * mmap_write_lock.
3286 */
3287 mas_set(&mas, vma->vm_end);
3288 do {
3289 if (vma->vm_flags & VM_ACCOUNT)
3290 nr_accounted += vma_pages(vma);
3291 remove_vma(vma, true);
3292 count++;
3293 cond_resched();
3294 vma = mas_find(&mas, ULONG_MAX);
3295 } while (vma && likely(!xa_is_zero(vma)));
3296
3297 BUG_ON(count != mm->map_count);
3298
3299 trace_exit_mmap(mm);
3300destroy:
3301 __mt_destroy(&mm->mm_mt);
3302 mmap_write_unlock(mm);
3303 vm_unacct_memory(nr_accounted);
3304}
3305
3306/* Insert vm structure into process list sorted by address
3307 * and into the inode's i_mmap tree. If vm_file is non-NULL
3308 * then i_mmap_rwsem is taken here.
3309 */
3310int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3311{
3312 unsigned long charged = vma_pages(vma);
3313
3314
3315 if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3316 return -ENOMEM;
3317
3318 if ((vma->vm_flags & VM_ACCOUNT) &&
3319 security_vm_enough_memory_mm(mm, charged))
3320 return -ENOMEM;
3321
3322 /*
3323 * The vm_pgoff of a purely anonymous vma should be irrelevant
3324 * until its first write fault, when page's anon_vma and index
3325 * are set. But now set the vm_pgoff it will almost certainly
3326 * end up with (unless mremap moves it elsewhere before that
3327 * first wfault), so /proc/pid/maps tells a consistent story.
3328 *
3329 * By setting it to reflect the virtual start address of the
3330 * vma, merges and splits can happen in a seamless way, just
3331 * using the existing file pgoff checks and manipulations.
3332 * Similarly in do_mmap and in do_brk_flags.
3333 */
3334 if (vma_is_anonymous(vma)) {
3335 BUG_ON(vma->anon_vma);
3336 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3337 }
3338
3339 if (vma_link(mm, vma)) {
3340 if (vma->vm_flags & VM_ACCOUNT)
3341 vm_unacct_memory(charged);
3342 return -ENOMEM;
3343 }
3344
3345 return 0;
3346}
3347
3348/*
3349 * Copy the vma structure to a new location in the same mm,
3350 * prior to moving page table entries, to effect an mremap move.
3351 */
3352struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3353 unsigned long addr, unsigned long len, pgoff_t pgoff,
3354 bool *need_rmap_locks)
3355{
3356 struct vm_area_struct *vma = *vmap;
3357 unsigned long vma_start = vma->vm_start;
3358 struct mm_struct *mm = vma->vm_mm;
3359 struct vm_area_struct *new_vma, *prev;
3360 bool faulted_in_anon_vma = true;
3361 VMA_ITERATOR(vmi, mm, addr);
3362
3363 /*
3364 * If anonymous vma has not yet been faulted, update new pgoff
3365 * to match new location, to increase its chance of merging.
3366 */
3367 if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3368 pgoff = addr >> PAGE_SHIFT;
3369 faulted_in_anon_vma = false;
3370 }
3371
3372 new_vma = find_vma_prev(mm, addr, &prev);
3373 if (new_vma && new_vma->vm_start < addr + len)
3374 return NULL; /* should never get here */
3375
3376 new_vma = vma_merge_new_vma(&vmi, prev, vma, addr, addr + len, pgoff);
3377 if (new_vma) {
3378 /*
3379 * Source vma may have been merged into new_vma
3380 */
3381 if (unlikely(vma_start >= new_vma->vm_start &&
3382 vma_start < new_vma->vm_end)) {
3383 /*
3384 * The only way we can get a vma_merge with
3385 * self during an mremap is if the vma hasn't
3386 * been faulted in yet and we were allowed to
3387 * reset the dst vma->vm_pgoff to the
3388 * destination address of the mremap to allow
3389 * the merge to happen. mremap must change the
3390 * vm_pgoff linearity between src and dst vmas
3391 * (in turn preventing a vma_merge) to be
3392 * safe. It is only safe to keep the vm_pgoff
3393 * linear if there are no pages mapped yet.
3394 */
3395 VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3396 *vmap = vma = new_vma;
3397 }
3398 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3399 } else {
3400 new_vma = vm_area_dup(vma);
3401 if (!new_vma)
3402 goto out;
3403 vma_set_range(new_vma, addr, addr + len, pgoff);
3404 if (vma_dup_policy(vma, new_vma))
3405 goto out_free_vma;
3406 if (anon_vma_clone(new_vma, vma))
3407 goto out_free_mempol;
3408 if (new_vma->vm_file)
3409 get_file(new_vma->vm_file);
3410 if (new_vma->vm_ops && new_vma->vm_ops->open)
3411 new_vma->vm_ops->open(new_vma);
3412 if (vma_link(mm, new_vma))
3413 goto out_vma_link;
3414 *need_rmap_locks = false;
3415 }
3416 return new_vma;
3417
3418out_vma_link:
3419 if (new_vma->vm_ops && new_vma->vm_ops->close)
3420 new_vma->vm_ops->close(new_vma);
3421
3422 if (new_vma->vm_file)
3423 fput(new_vma->vm_file);
3424
3425 unlink_anon_vmas(new_vma);
3426out_free_mempol:
3427 mpol_put(vma_policy(new_vma));
3428out_free_vma:
3429 vm_area_free(new_vma);
3430out:
3431 return NULL;
3432}
3433
3434/*
3435 * Return true if the calling process may expand its vm space by the passed
3436 * number of pages
3437 */
3438bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3439{
3440 if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3441 return false;
3442
3443 if (is_data_mapping(flags) &&
3444 mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3445 /* Workaround for Valgrind */
3446 if (rlimit(RLIMIT_DATA) == 0 &&
3447 mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3448 return true;
3449
3450 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3451 current->comm, current->pid,
3452 (mm->data_vm + npages) << PAGE_SHIFT,
3453 rlimit(RLIMIT_DATA),
3454 ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3455
3456 if (!ignore_rlimit_data)
3457 return false;
3458 }
3459
3460 return true;
3461}
3462
3463void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3464{
3465 WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3466
3467 if (is_exec_mapping(flags))
3468 mm->exec_vm += npages;
3469 else if (is_stack_mapping(flags))
3470 mm->stack_vm += npages;
3471 else if (is_data_mapping(flags))
3472 mm->data_vm += npages;
3473}
3474
3475static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3476
3477/*
3478 * Having a close hook prevents vma merging regardless of flags.
3479 */
3480static void special_mapping_close(struct vm_area_struct *vma)
3481{
3482}
3483
3484static const char *special_mapping_name(struct vm_area_struct *vma)
3485{
3486 return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3487}
3488
3489static int special_mapping_mremap(struct vm_area_struct *new_vma)
3490{
3491 struct vm_special_mapping *sm = new_vma->vm_private_data;
3492
3493 if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3494 return -EFAULT;
3495
3496 if (sm->mremap)
3497 return sm->mremap(sm, new_vma);
3498
3499 return 0;
3500}
3501
3502static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3503{
3504 /*
3505 * Forbid splitting special mappings - kernel has expectations over
3506 * the number of pages in mapping. Together with VM_DONTEXPAND
3507 * the size of vma should stay the same over the special mapping's
3508 * lifetime.
3509 */
3510 return -EINVAL;
3511}
3512
3513static const struct vm_operations_struct special_mapping_vmops = {
3514 .close = special_mapping_close,
3515 .fault = special_mapping_fault,
3516 .mremap = special_mapping_mremap,
3517 .name = special_mapping_name,
3518 /* vDSO code relies that VVAR can't be accessed remotely */
3519 .access = NULL,
3520 .may_split = special_mapping_split,
3521};
3522
3523static const struct vm_operations_struct legacy_special_mapping_vmops = {
3524 .close = special_mapping_close,
3525 .fault = special_mapping_fault,
3526};
3527
3528static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3529{
3530 struct vm_area_struct *vma = vmf->vma;
3531 pgoff_t pgoff;
3532 struct page **pages;
3533
3534 if (vma->vm_ops == &legacy_special_mapping_vmops) {
3535 pages = vma->vm_private_data;
3536 } else {
3537 struct vm_special_mapping *sm = vma->vm_private_data;
3538
3539 if (sm->fault)
3540 return sm->fault(sm, vmf->vma, vmf);
3541
3542 pages = sm->pages;
3543 }
3544
3545 for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3546 pgoff--;
3547
3548 if (*pages) {
3549 struct page *page = *pages;
3550 get_page(page);
3551 vmf->page = page;
3552 return 0;
3553 }
3554
3555 return VM_FAULT_SIGBUS;
3556}
3557
3558static struct vm_area_struct *__install_special_mapping(
3559 struct mm_struct *mm,
3560 unsigned long addr, unsigned long len,
3561 unsigned long vm_flags, void *priv,
3562 const struct vm_operations_struct *ops)
3563{
3564 int ret;
3565 struct vm_area_struct *vma;
3566
3567 vma = vm_area_alloc(mm);
3568 if (unlikely(vma == NULL))
3569 return ERR_PTR(-ENOMEM);
3570
3571 vma_set_range(vma, addr, addr + len, 0);
3572 vm_flags_init(vma, (vm_flags | mm->def_flags |
3573 VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
3574 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3575
3576 vma->vm_ops = ops;
3577 vma->vm_private_data = priv;
3578
3579 ret = insert_vm_struct(mm, vma);
3580 if (ret)
3581 goto out;
3582
3583 vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3584
3585 perf_event_mmap(vma);
3586
3587 return vma;
3588
3589out:
3590 vm_area_free(vma);
3591 return ERR_PTR(ret);
3592}
3593
3594bool vma_is_special_mapping(const struct vm_area_struct *vma,
3595 const struct vm_special_mapping *sm)
3596{
3597 return vma->vm_private_data == sm &&
3598 (vma->vm_ops == &special_mapping_vmops ||
3599 vma->vm_ops == &legacy_special_mapping_vmops);
3600}
3601
3602/*
3603 * Called with mm->mmap_lock held for writing.
3604 * Insert a new vma covering the given region, with the given flags.
3605 * Its pages are supplied by the given array of struct page *.
3606 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3607 * The region past the last page supplied will always produce SIGBUS.
3608 * The array pointer and the pages it points to are assumed to stay alive
3609 * for as long as this mapping might exist.
3610 */
3611struct vm_area_struct *_install_special_mapping(
3612 struct mm_struct *mm,
3613 unsigned long addr, unsigned long len,
3614 unsigned long vm_flags, const struct vm_special_mapping *spec)
3615{
3616 return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3617 &special_mapping_vmops);
3618}
3619
3620int install_special_mapping(struct mm_struct *mm,
3621 unsigned long addr, unsigned long len,
3622 unsigned long vm_flags, struct page **pages)
3623{
3624 struct vm_area_struct *vma = __install_special_mapping(
3625 mm, addr, len, vm_flags, (void *)pages,
3626 &legacy_special_mapping_vmops);
3627
3628 return PTR_ERR_OR_ZERO(vma);
3629}
3630
3631static DEFINE_MUTEX(mm_all_locks_mutex);
3632
3633static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3634{
3635 if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3636 /*
3637 * The LSB of head.next can't change from under us
3638 * because we hold the mm_all_locks_mutex.
3639 */
3640 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3641 /*
3642 * We can safely modify head.next after taking the
3643 * anon_vma->root->rwsem. If some other vma in this mm shares
3644 * the same anon_vma we won't take it again.
3645 *
3646 * No need of atomic instructions here, head.next
3647 * can't change from under us thanks to the
3648 * anon_vma->root->rwsem.
3649 */
3650 if (__test_and_set_bit(0, (unsigned long *)
3651 &anon_vma->root->rb_root.rb_root.rb_node))
3652 BUG();
3653 }
3654}
3655
3656static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3657{
3658 if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3659 /*
3660 * AS_MM_ALL_LOCKS can't change from under us because
3661 * we hold the mm_all_locks_mutex.
3662 *
3663 * Operations on ->flags have to be atomic because
3664 * even if AS_MM_ALL_LOCKS is stable thanks to the
3665 * mm_all_locks_mutex, there may be other cpus
3666 * changing other bitflags in parallel to us.
3667 */
3668 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3669 BUG();
3670 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3671 }
3672}
3673
3674/*
3675 * This operation locks against the VM for all pte/vma/mm related
3676 * operations that could ever happen on a certain mm. This includes
3677 * vmtruncate, try_to_unmap, and all page faults.
3678 *
3679 * The caller must take the mmap_lock in write mode before calling
3680 * mm_take_all_locks(). The caller isn't allowed to release the
3681 * mmap_lock until mm_drop_all_locks() returns.
3682 *
3683 * mmap_lock in write mode is required in order to block all operations
3684 * that could modify pagetables and free pages without need of
3685 * altering the vma layout. It's also needed in write mode to avoid new
3686 * anon_vmas to be associated with existing vmas.
3687 *
3688 * A single task can't take more than one mm_take_all_locks() in a row
3689 * or it would deadlock.
3690 *
3691 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3692 * mapping->flags avoid to take the same lock twice, if more than one
3693 * vma in this mm is backed by the same anon_vma or address_space.
3694 *
3695 * We take locks in following order, accordingly to comment at beginning
3696 * of mm/rmap.c:
3697 * - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3698 * hugetlb mapping);
3699 * - all vmas marked locked
3700 * - all i_mmap_rwsem locks;
3701 * - all anon_vma->rwseml
3702 *
3703 * We can take all locks within these types randomly because the VM code
3704 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3705 * mm_all_locks_mutex.
3706 *
3707 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3708 * that may have to take thousand of locks.
3709 *
3710 * mm_take_all_locks() can fail if it's interrupted by signals.
3711 */
3712int mm_take_all_locks(struct mm_struct *mm)
3713{
3714 struct vm_area_struct *vma;
3715 struct anon_vma_chain *avc;
3716 MA_STATE(mas, &mm->mm_mt, 0, 0);
3717
3718 mmap_assert_write_locked(mm);
3719
3720 mutex_lock(&mm_all_locks_mutex);
3721
3722 /*
3723 * vma_start_write() does not have a complement in mm_drop_all_locks()
3724 * because vma_start_write() is always asymmetrical; it marks a VMA as
3725 * being written to until mmap_write_unlock() or mmap_write_downgrade()
3726 * is reached.
3727 */
3728 mas_for_each(&mas, vma, ULONG_MAX) {
3729 if (signal_pending(current))
3730 goto out_unlock;
3731 vma_start_write(vma);
3732 }
3733
3734 mas_set(&mas, 0);
3735 mas_for_each(&mas, vma, ULONG_MAX) {
3736 if (signal_pending(current))
3737 goto out_unlock;
3738 if (vma->vm_file && vma->vm_file->f_mapping &&
3739 is_vm_hugetlb_page(vma))
3740 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3741 }
3742
3743 mas_set(&mas, 0);
3744 mas_for_each(&mas, vma, ULONG_MAX) {
3745 if (signal_pending(current))
3746 goto out_unlock;
3747 if (vma->vm_file && vma->vm_file->f_mapping &&
3748 !is_vm_hugetlb_page(vma))
3749 vm_lock_mapping(mm, vma->vm_file->f_mapping);
3750 }
3751
3752 mas_set(&mas, 0);
3753 mas_for_each(&mas, vma, ULONG_MAX) {
3754 if (signal_pending(current))
3755 goto out_unlock;
3756 if (vma->anon_vma)
3757 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3758 vm_lock_anon_vma(mm, avc->anon_vma);
3759 }
3760
3761 return 0;
3762
3763out_unlock:
3764 mm_drop_all_locks(mm);
3765 return -EINTR;
3766}
3767
3768static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3769{
3770 if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3771 /*
3772 * The LSB of head.next can't change to 0 from under
3773 * us because we hold the mm_all_locks_mutex.
3774 *
3775 * We must however clear the bitflag before unlocking
3776 * the vma so the users using the anon_vma->rb_root will
3777 * never see our bitflag.
3778 *
3779 * No need of atomic instructions here, head.next
3780 * can't change from under us until we release the
3781 * anon_vma->root->rwsem.
3782 */
3783 if (!__test_and_clear_bit(0, (unsigned long *)
3784 &anon_vma->root->rb_root.rb_root.rb_node))
3785 BUG();
3786 anon_vma_unlock_write(anon_vma);
3787 }
3788}
3789
3790static void vm_unlock_mapping(struct address_space *mapping)
3791{
3792 if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3793 /*
3794 * AS_MM_ALL_LOCKS can't change to 0 from under us
3795 * because we hold the mm_all_locks_mutex.
3796 */
3797 i_mmap_unlock_write(mapping);
3798 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3799 &mapping->flags))
3800 BUG();
3801 }
3802}
3803
3804/*
3805 * The mmap_lock cannot be released by the caller until
3806 * mm_drop_all_locks() returns.
3807 */
3808void mm_drop_all_locks(struct mm_struct *mm)
3809{
3810 struct vm_area_struct *vma;
3811 struct anon_vma_chain *avc;
3812 MA_STATE(mas, &mm->mm_mt, 0, 0);
3813
3814 mmap_assert_write_locked(mm);
3815 BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3816
3817 mas_for_each(&mas, vma, ULONG_MAX) {
3818 if (vma->anon_vma)
3819 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3820 vm_unlock_anon_vma(avc->anon_vma);
3821 if (vma->vm_file && vma->vm_file->f_mapping)
3822 vm_unlock_mapping(vma->vm_file->f_mapping);
3823 }
3824
3825 mutex_unlock(&mm_all_locks_mutex);
3826}
3827
3828/*
3829 * initialise the percpu counter for VM
3830 */
3831void __init mmap_init(void)
3832{
3833 int ret;
3834
3835 ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3836 VM_BUG_ON(ret);
3837}
3838
3839/*
3840 * Initialise sysctl_user_reserve_kbytes.
3841 *
3842 * This is intended to prevent a user from starting a single memory hogging
3843 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3844 * mode.
3845 *
3846 * The default value is min(3% of free memory, 128MB)
3847 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3848 */
3849static int init_user_reserve(void)
3850{
3851 unsigned long free_kbytes;
3852
3853 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3854
3855 sysctl_user_reserve_kbytes = min(free_kbytes / 32, SZ_128K);
3856 return 0;
3857}
3858subsys_initcall(init_user_reserve);
3859
3860/*
3861 * Initialise sysctl_admin_reserve_kbytes.
3862 *
3863 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3864 * to log in and kill a memory hogging process.
3865 *
3866 * Systems with more than 256MB will reserve 8MB, enough to recover
3867 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3868 * only reserve 3% of free pages by default.
3869 */
3870static int init_admin_reserve(void)
3871{
3872 unsigned long free_kbytes;
3873
3874 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3875
3876 sysctl_admin_reserve_kbytes = min(free_kbytes / 32, SZ_8K);
3877 return 0;
3878}
3879subsys_initcall(init_admin_reserve);
3880
3881/*
3882 * Reinititalise user and admin reserves if memory is added or removed.
3883 *
3884 * The default user reserve max is 128MB, and the default max for the
3885 * admin reserve is 8MB. These are usually, but not always, enough to
3886 * enable recovery from a memory hogging process using login/sshd, a shell,
3887 * and tools like top. It may make sense to increase or even disable the
3888 * reserve depending on the existence of swap or variations in the recovery
3889 * tools. So, the admin may have changed them.
3890 *
3891 * If memory is added and the reserves have been eliminated or increased above
3892 * the default max, then we'll trust the admin.
3893 *
3894 * If memory is removed and there isn't enough free memory, then we
3895 * need to reset the reserves.
3896 *
3897 * Otherwise keep the reserve set by the admin.
3898 */
3899static int reserve_mem_notifier(struct notifier_block *nb,
3900 unsigned long action, void *data)
3901{
3902 unsigned long tmp, free_kbytes;
3903
3904 switch (action) {
3905 case MEM_ONLINE:
3906 /* Default max is 128MB. Leave alone if modified by operator. */
3907 tmp = sysctl_user_reserve_kbytes;
3908 if (tmp > 0 && tmp < SZ_128K)
3909 init_user_reserve();
3910
3911 /* Default max is 8MB. Leave alone if modified by operator. */
3912 tmp = sysctl_admin_reserve_kbytes;
3913 if (tmp > 0 && tmp < SZ_8K)
3914 init_admin_reserve();
3915
3916 break;
3917 case MEM_OFFLINE:
3918 free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3919
3920 if (sysctl_user_reserve_kbytes > free_kbytes) {
3921 init_user_reserve();
3922 pr_info("vm.user_reserve_kbytes reset to %lu\n",
3923 sysctl_user_reserve_kbytes);
3924 }
3925
3926 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3927 init_admin_reserve();
3928 pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3929 sysctl_admin_reserve_kbytes);
3930 }
3931 break;
3932 default:
3933 break;
3934 }
3935 return NOTIFY_OK;
3936}
3937
3938static int __meminit init_reserve_notifier(void)
3939{
3940 if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3941 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3942
3943 return 0;
3944}
3945subsys_initcall(init_reserve_notifier);