Mercurial > dropbear

comparison FUZZER-NOTES.md @ 1590:68d5d8e84a92

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
notes on the fuzzer
author Matt Johnston <matt@ucc.asn.au>
date Mon, 05 Mar 2018 14:07:11 +0800
parents
children b794d277c6da
comparison
equal deleted inserted replaced
1589:35af85194268 1590:68d5d8e84a92
1 # Fuzzing Dropbear
2
3 Dropbear is process-per-session so it assumes calling `dropbear_exit()`
4 is fine at any point to clean up. This makes fuzzing a bit trickier.
5 A few pieces of wrapping infrastructure are used to work around this.
6
7 The [libfuzzer](http://llvm.org/docs/LibFuzzer.html#fuzz-target) harness
8 expects a long running process to continually run a test function with
9 a string of crafted input. That process should not leak resources or exit.
10
11 ## longjmp
12
13 When dropbear runs in fuzz mode it sets up a
14 [`setjmp()`](http://man7.org/linux/man-pages/man3/setjmp.3.html) target prior
15 to launching the code to be fuzzed, and then [`dropbear_exit()`](dbutil.c#L125)
16 calls `longjmp()` back there. This avoids exiting though it doesn't free
17 memory or other resources.
18
19 ## malloc Wrapper
20
21 Dropbear normally uses a [`m_malloc()`](dbmalloc.c) function that is the same as `malloc()` but
22 exits if allocation fails. In fuzzing mode this is replaced with a tracking allocator
23 that stores all allocations in a linked list. After the `longjmp()` occurs the fuzzer target
24 calls [`m_malloc_free_epoch(1, 1)`](dbmalloc.c) to clean up any unreleased memory.
25
26 If the fuzz target runs to completion it calls `m_malloc_free_epoch(1, 0)` which will reset
27 the tracked allocations but will not free memory - that allows libfuzzer's leak checking
28 to detect leaks in normal operation.
29
30 ## File Descriptor Input
31
32 As a network process Dropbear reads and writes from a socket. The wrappers for
33 `read()`/`write()`/`select()` in [fuzz-wrapfd.c](fuzz-wrapfd.c) will read from the
34 fuzzer input that has been set up with `wrapfd_add()`. `write()` output is
35 currently discarded.
36 These also test error paths such as EINTR and short reads with certain probabilities.
37
38 This allows running the entire dropbear server process with network input provided by the
39 fuzzer, without many modifications to the main code. At the time of writing this
40 only runs the pre-authentication stages, though post-authentication could be run similarly.
41
42 ## Encryption and Randomness
43
44 When running in fuzzing mode Dropbear uses a [fixed seed](dbrandom.c#L185)
45 every time so that failures can be reproduced.
46
47 Since the fuzzer cannot generate valid encrypted input the packet decryption and
48 message authentication calls are disabled, see (packet.c)[packet.c].
49 MAC failures are set to occur with a low probability to test that error path.
50
51 ## Fuzzers
52
53 Current fuzzers are
54
55 - fuzzer-preauth - the fuzzer input is treated as a stream of session input. This will
56 test key exchange, packet ordering, authentication attempts etc.
57
58 - fuzzer-preauth_nomaths - the same as fuzzer-preauth but with asymmetric crypto
59 routines replaced with dummies for faster runtime. corpora are shared
60 between fuzzers by [oss-fuzz](https://github.com/google/oss-fuzz) so this
61 will help fuzzer-preauth too.
62
63 - fuzzer-verify - read a key and signature from fuzzer input and verify that signature.
64 It would not be expected to pass, though some keys with bad parameters are
65 able to validate with a trivial signature - extra checks are added for that.
66
67 - fuzzer-pubkey - test parsing of an `authorized_keys` line.
68
69 - fuzzer-kexdh - test Diffie-Hellman key exchange where the fuzz input is the
70 public key that would be received over the network.
71
72 - fuzzer-kexecdh - test Elliptic Curve Diffie-Hellman key exchange like fuzzer-kexdh