Commit | Line | Data |
---|---|---|
f84a56f7 MR |
1 | This document explains potential effects of speculation, and how undesirable |
2 | effects can be mitigated portably using common APIs. | |
3 | ||
4 | =========== | |
5 | Speculation | |
6 | =========== | |
7 | ||
8 | To improve performance and minimize average latencies, many contemporary CPUs | |
9 | employ speculative execution techniques such as branch prediction, performing | |
10 | work which may be discarded at a later stage. | |
11 | ||
12 | Typically speculative execution cannot be observed from architectural state, | |
13 | such as the contents of registers. However, in some cases it is possible to | |
14 | observe its impact on microarchitectural state, such as the presence or | |
15 | absence of data in caches. Such state may form side-channels which can be | |
16 | observed to extract secret information. | |
17 | ||
18 | For example, in the presence of branch prediction, it is possible for bounds | |
19 | checks to be ignored by code which is speculatively executed. Consider the | |
20 | following code: | |
21 | ||
22 | int load_array(int *array, unsigned int index) | |
23 | { | |
24 | if (index >= MAX_ARRAY_ELEMS) | |
25 | return 0; | |
26 | else | |
27 | return array[index]; | |
28 | } | |
29 | ||
30 | Which, on arm64, may be compiled to an assembly sequence such as: | |
31 | ||
32 | CMP <index>, #MAX_ARRAY_ELEMS | |
33 | B.LT less | |
34 | MOV <returnval>, #0 | |
35 | RET | |
36 | less: | |
37 | LDR <returnval>, [<array>, <index>] | |
38 | RET | |
39 | ||
40 | It is possible that a CPU mis-predicts the conditional branch, and | |
41 | speculatively loads array[index], even if index >= MAX_ARRAY_ELEMS. This | |
42 | value will subsequently be discarded, but the speculated load may affect | |
43 | microarchitectural state which can be subsequently measured. | |
44 | ||
45 | More complex sequences involving multiple dependent memory accesses may | |
46 | result in sensitive information being leaked. Consider the following | |
47 | code, building on the prior example: | |
48 | ||
49 | int load_dependent_arrays(int *arr1, int *arr2, int index) | |
50 | { | |
51 | int val1, val2, | |
52 | ||
53 | val1 = load_array(arr1, index); | |
54 | val2 = load_array(arr2, val1); | |
55 | ||
56 | return val2; | |
57 | } | |
58 | ||
59 | Under speculation, the first call to load_array() may return the value | |
60 | of an out-of-bounds address, while the second call will influence | |
61 | microarchitectural state dependent on this value. This may provide an | |
62 | arbitrary read primitive. | |
63 | ||
64 | ==================================== | |
65 | Mitigating speculation side-channels | |
66 | ==================================== | |
67 | ||
68 | The kernel provides a generic API to ensure that bounds checks are | |
69 | respected even under speculation. Architectures which are affected by | |
70 | speculation-based side-channels are expected to implement these | |
71 | primitives. | |
72 | ||
73 | The array_index_nospec() helper in <linux/nospec.h> can be used to | |
74 | prevent information from being leaked via side-channels. | |
75 | ||
76 | A call to array_index_nospec(index, size) returns a sanitized index | |
77 | value that is bounded to [0, size) even under cpu speculation | |
78 | conditions. | |
79 | ||
80 | This can be used to protect the earlier load_array() example: | |
81 | ||
82 | int load_array(int *array, unsigned int index) | |
83 | { | |
84 | if (index >= MAX_ARRAY_ELEMS) | |
85 | return 0; | |
86 | else { | |
87 | index = array_index_nospec(index, MAX_ARRAY_ELEMS); | |
88 | return array[index]; | |
89 | } | |
90 | } |