view FUZZER-NOTES.md @ 1788:1fc0012b9c38

Fix handling of replies to global requests (#112) The current code assumes that all global requests want / need a reply. This isn't always true and the request itself indicates if it wants a reply or not. It causes a specific problem with [email protected] messages. These are sent by OpenSSH after authentication to inform the client of potential other host keys for the host. This can be used to add a new type of host key or to rotate host keys. The initial information message from the server is sent as a global request, but with want_reply set to false. This means that the server doesn't expect an answer to this message. Instead the client needs to send a prove request as a reply if it wants to receive proof of ownership for the host keys. The bug doesn't cause any current problems with due to how OpenSSH treats receiving the failure message. It instead treats it as a keepalive message and further ignores it. Arguably this is a protocol violation though of Dropbear and it is only accidental that it doesn't cause a problem with OpenSSH. The bug was found when adding host keys support to libssh, which is more strict protocol wise and treats the unexpected failure message an error, also see https://gitlab.com/libssh/libssh-mirror/-/merge_requests/145 for more information. The fix here is to honor the want_reply flag in the global request and to only send a reply if the other side expects a reply.
author Dirkjan Bussink <d.bussink@gmail.com>
date Thu, 10 Dec 2020 16:13:13 +0100
parents d32bcb5c557d
children
line wrap: on
line source

# Fuzzing Dropbear

Dropbear is process-per-session so it assumes calling `dropbear_exit()`
is fine at any point to clean up. This makes fuzzing a bit trickier. 
A few pieces of wrapping infrastructure are used to work around this.

The [libfuzzer](http://llvm.org/docs/LibFuzzer.html#fuzz-target) harness
expects a long running process to continually run a test function with 
a string of crafted input. That process should not leak resources or exit.

## longjmp

When dropbear runs in fuzz mode it sets up a 
[`setjmp()`](http://man7.org/linux/man-pages/man3/setjmp.3.html) target prior 
to launching the code to be fuzzed, and then [`dropbear_exit()`](dbutil.c#L125)
calls `longjmp()` back there. This avoids exiting though it doesn't free 
memory or other resources.

## malloc Wrapper

Dropbear normally uses a [`m_malloc()`](dbmalloc.c) function that is the same as `malloc()` but
exits if allocation fails. In fuzzing mode this is replaced with a tracking allocator
that stores all allocations in a linked list. After the `longjmp()` occurs the fuzzer target
calls [`m_malloc_free_epoch(1, 1)`](dbmalloc.c) to clean up any unreleased memory.

If the fuzz target runs to completion it calls `m_malloc_free_epoch(1, 0)` which will reset 
the tracked allocations but will not free memory - that allows libfuzzer's leak checking
to detect leaks in normal operation.

## File Descriptor Input

As a network process Dropbear reads and writes from a socket. The wrappers for
`read()`/`write()`/`select()` in [fuzz-wrapfd.c](fuzz-wrapfd.c) will read from the
fuzzer input that has been set up with `wrapfd_add()`. `write()` output is
currently discarded.
These also test error paths such as EINTR and short reads with certain probabilities.

This allows running the entire dropbear server process with network input provided by the
fuzzer, without many modifications to the main code. At the time of writing this 
only runs the pre-authentication stages, though post-authentication could be run similarly.

## Encryption and Randomness

When running in fuzzing mode Dropbear uses a [fixed seed](dbrandom.c#L185)
every time so that failures can be reproduced. 

Since the fuzzer cannot generate valid encrypted input the packet decryption and
message authentication calls are disabled, see [packet.c](packet.c). 
MAC failures are set to occur with a low probability to test that error path.

## Fuzzers

Current fuzzers are

- [fuzzer-preauth](fuzzer-preauth.c) - the fuzzer input is treated as a stream of session input. This will
  test key exchange, packet ordering, authentication attempts etc.

- [fuzzer-preauth_nomaths](fuzzer-preauth_nomaths.c) - the same as fuzzer-preauth but with asymmetric crypto
  routines replaced with dummies for faster runtime. corpora are shared 
  between fuzzers by [oss-fuzz](https://github.com/google/oss-fuzz) so this 
  will help fuzzer-preauth too.

- [fuzzer-verify](fuzzer-verify.c) - read a key and signature from fuzzer input and verify that signature. 
  It would not be expected to pass, though some keys with bad parameters are 
  able to validate with a trivial signature - extra checks are added for that.

- [fuzzer-pubkey](fuzzer-pubkey.c) - test parsing of an `authorized_keys` line.

- [fuzzer-kexdh](fuzzer-kexdh.c) - test Diffie-Hellman key exchange where the fuzz input is the 
  ephemeral public key that would be received over the network. This is testing `mp_expt_mod()`
  and and other libtommath routines.

- [fuzzer-kexecdh](fuzzer-kexecdh.c) - test Elliptic Curve Diffie-Hellman key exchange like fuzzer-kexdh.
  This is testing libtommath ECC routines.

- [fuzzer-kexcurve25519](fuzzer-kexcurve25519.c) - test Curve25519 Elliptic Curve Diffie-Hellman key exchange
  like fuzzer-kexecdh. This is testing `dropbear_curve25519_scalarmult()` and other libtommath routines.