1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Implement mseal() syscall.
4 *
5 * Copyright (c) 2023,2024 Google, Inc.
6 *
7 * Author: Jeff Xu <jeffxu@chromium.org>
8 */
9
10 #include <linux/mempolicy.h>
11 #include <linux/mman.h>
12 #include <linux/mm.h>
13 #include <linux/mm_inline.h>
14 #include <linux/syscalls.h>
15 #include <linux/sched.h>
16 #include "internal.h"
17
18 /*
19 * mseal() disallows an input range which contain unmapped ranges (VMA holes).
20 *
21 * It disallows unmapped regions from start to end whether they exist at the
22 * start, in the middle, or at the end of the range, or any combination thereof.
23 *
24 * This is because after sealng a range, there's nothing to stop memory mapping
25 * of ranges in the remaining gaps later, meaning that the user might then
26 * wrongly consider the entirety of the mseal()'d range to be sealed when it
27 * in fact isn't.
28 */
29
30 /*
31 * Does the [start, end) range contain any unmapped memory?
32 *
33 * We ensure that:
34 * - start is part of a valid VMA.
35 * - end is part of a valid VMA.
36 * - no gap (unallocated memory) exists between start and end.
37 */
range_contains_unmapped(struct mm_struct * mm,unsigned long start,unsigned long end)38 static bool range_contains_unmapped(struct mm_struct *mm,
39 unsigned long start, unsigned long end)
40 {
41 struct vm_area_struct *vma;
42 unsigned long prev_end = start;
43 VMA_ITERATOR(vmi, current->mm, start);
44
45 for_each_vma_range(vmi, vma, end) {
46 if (vma->vm_start > prev_end)
47 return true;
48
49 prev_end = vma->vm_end;
50 }
51
52 return prev_end < end;
53 }
54
mseal_apply(struct mm_struct * mm,unsigned long start,unsigned long end)55 static int mseal_apply(struct mm_struct *mm,
56 unsigned long start, unsigned long end)
57 {
58 struct vm_area_struct *vma, *prev;
59 unsigned long curr_start = start;
60 VMA_ITERATOR(vmi, mm, start);
61
62 /* We know there are no gaps so this will be non-NULL. */
63 vma = vma_iter_load(&vmi);
64 prev = vma_prev(&vmi);
65 if (start > vma->vm_start)
66 prev = vma;
67
68 for_each_vma_range(vmi, vma, end) {
69 unsigned long curr_end = MIN(vma->vm_end, end);
70
71 if (!(vma->vm_flags & VM_SEALED)) {
72 vma = vma_modify_flags(&vmi, prev, vma,
73 curr_start, curr_end,
74 vma->vm_flags | VM_SEALED);
75 if (IS_ERR(vma))
76 return PTR_ERR(vma);
77 vm_flags_set(vma, VM_SEALED);
78 }
79
80 prev = vma;
81 curr_start = curr_end;
82 }
83
84 return 0;
85 }
86
87 /*
88 * mseal(2) seals the VM's meta data from
89 * selected syscalls.
90 *
91 * addr/len: VM address range.
92 *
93 * The address range by addr/len must meet:
94 * start (addr) must be in a valid VMA.
95 * end (addr + len) must be in a valid VMA.
96 * no gap (unallocated memory) between start and end.
97 * start (addr) must be page aligned.
98 *
99 * len: len will be page aligned implicitly.
100 *
101 * Below VMA operations are blocked after sealing.
102 * 1> Unmapping, moving to another location, and shrinking
103 * the size, via munmap() and mremap(), can leave an empty
104 * space, therefore can be replaced with a VMA with a new
105 * set of attributes.
106 * 2> Moving or expanding a different vma into the current location,
107 * via mremap().
108 * 3> Modifying a VMA via mmap(MAP_FIXED).
109 * 4> Size expansion, via mremap(), does not appear to pose any
110 * specific risks to sealed VMAs. It is included anyway because
111 * the use case is unclear. In any case, users can rely on
112 * merging to expand a sealed VMA.
113 * 5> mprotect and pkey_mprotect.
114 * 6> Some destructive madvice() behavior (e.g. MADV_DONTNEED)
115 * for anonymous memory, when users don't have write permission to the
116 * memory. Those behaviors can alter region contents by discarding pages,
117 * effectively a memset(0) for anonymous memory.
118 *
119 * flags: reserved.
120 *
121 * return values:
122 * zero: success.
123 * -EINVAL:
124 * invalid input flags.
125 * start address is not page aligned.
126 * Address arange (start + len) overflow.
127 * -ENOMEM:
128 * addr is not a valid address (not allocated).
129 * end (start + len) is not a valid address.
130 * a gap (unallocated memory) between start and end.
131 * -EPERM:
132 * - In 32 bit architecture, sealing is not supported.
133 * Note:
134 * user can call mseal(2) multiple times, adding a seal on an
135 * already sealed memory is a no-action (no error).
136 *
137 * unseal() is not supported.
138 */
do_mseal(unsigned long start,size_t len_in,unsigned long flags)139 int do_mseal(unsigned long start, size_t len_in, unsigned long flags)
140 {
141 size_t len;
142 int ret = 0;
143 unsigned long end;
144 struct mm_struct *mm = current->mm;
145
146 /* Verify flags not set. */
147 if (flags)
148 return -EINVAL;
149
150 start = untagged_addr(start);
151 if (!PAGE_ALIGNED(start))
152 return -EINVAL;
153
154 len = PAGE_ALIGN(len_in);
155 /* Check to see whether len was rounded up from small -ve to zero. */
156 if (len_in && !len)
157 return -EINVAL;
158
159 end = start + len;
160 if (end < start)
161 return -EINVAL;
162
163 if (end == start)
164 return 0;
165
166 if (mmap_write_lock_killable(mm))
167 return -EINTR;
168
169 if (range_contains_unmapped(mm, start, end)) {
170 ret = -ENOMEM;
171 goto out;
172 }
173
174 /*
175 * Second pass, this should success, unless there are errors
176 * from vma_modify_flags, e.g. merge/split error, or process
177 * reaching the max supported VMAs, however, those cases shall
178 * be rare.
179 */
180 ret = mseal_apply(mm, start, end);
181
182 out:
183 mmap_write_unlock(mm);
184 return ret;
185 }
186
SYSCALL_DEFINE3(mseal,unsigned long,start,size_t,len,unsigned long,flags)187 SYSCALL_DEFINE3(mseal, unsigned long, start, size_t, len, unsigned long,
188 flags)
189 {
190 return do_mseal(start, len, flags);
191 }
192