test-fuzz
is a Cargo subcommand and a collection of Rust macros to automate certain tasks related to fuzzing with afl.rs
, including:
- generating a fuzzing corpus
- implementing a fuzzing harness
test-fuzz
accomplishes these (in part) using Rust's testing facilities. For example, to generate a fuzzing corpus, test-fuzz
records a target's arguments each time it is called during an invocation of cargo test
. Similarly, test-fuzz
implements a fuzzing harness as an additional test in a cargo-test
-generated binary. This tight integration with Rust's testing facilities is what motivates the name test
-fuzz
.
Contents
- Installation
- Overview
- Components
test-fuzz
package features- Auto-generated corpus files
- Environment variables
- Limitations
- Tips and tricks
- [Semantic versioning policy]
- License
Install cargo-test-fuzz
and afl.rs
with the following command:
cargo install cargo-test-fuzz cargo-afl
Fuzzing with test-fuzz
is essentially three steps:*
-
Identify a fuzz target:
- Add the following
dependencies
to the target crate'sCargo.toml
file:serde = "*" test-fuzz = "*"
- Precede the target function with the
test_fuzz
macro:#[test_fuzz::test_fuzz] fn foo(...) { ... }
- Add the following
-
Generate a corpus by running
cargo test
:cargo test
-
Fuzz your target by running
cargo test-fuzz
:cargo test-fuzz foo
* An additional, preliminary step may be necessary following a reboot:
cargo afl system-config
Note that the above command runs sudo
internally. Hence, you may be prompted to enter your password.
Preceding a function with the test_fuzz
macro indicates that the function is a fuzz target.
The primary effects of the test_fuzz
macro are:
- Add instrumentation to the target to serialize its arguments and write them to a corpus file each time the target is called. The instrumentation is guarded by
#[cfg(test)]
so that corpus files are generated only when running tests (however, seeenable_in_production
below). - Add a test to read and deserialize arguments from standard input and apply the target to them. The test checks an environment variable, set by
cargo test-fuzz
, so that the test does not block trying to read from standard input during a normal invocation ofcargo test
. The test is enclosed in a module to reduce the likelihood of a name collision. Currently, the name of the module istarget_fuzz
, wheretarget
is the name of the target (however, seerename
below).
Impose where_predicates
(e.g., trait bounds) on the struct used to serialize/deserialize arguments. This may be necessary, e.g., if a target's argument type is an associated type. For an example, see associated_type.rs in this repository.
Use parameters
as the target's type parameters when fuzzing. Example:
#[test_fuzz(generic_args = "String")]
fn foo<T: Clone + Debug + Serialize>(x: &T) {
...
}
Note: The target's arguments must be serializable for every instantiation of its type parameters. But the target's arguments are required to be deserializable only when the target is instantiated with parameters
.
Use parameters
as the target's Self
type parameters when fuzzing. Example:
#[test_fuzz_impl]
impl<T: Clone + Debug + Serialize> for Foo {
#[test_fuzz(impl_generic_args = "String")]
fn bar(&self, x: &T) {
...
}
}
Note: The target's arguments must be serializable for every instantiation of its Self
type parameters. But the target's arguments are required to be deserializable only when the target's Self
is instantiated with parameters
.
When serializing the target's arguments, convert values of type X
to type Y
using Y
's implementation of From<X>
, or of type &X
to type Y
using Y
's implementation of the non-standard trait test_fuzz::FromRef<X>
. When deserializing, convert those values back to type X
using Y
's implementation of the non-standard trait test_fuzz::Into<X>
.
That is, use of convert = "X, Y"
must be accompanied by certain implementations. If X
implements Clone
, then Y
may implement the following:
impl From<X> for Y {
fn from(x: X) -> Self {
...
}
}
If X
does not implement Clone
, then Y
must implement the following:
impl test_fuzz::FromRef<X> for Y {
fn from_ref(x: &X) -> Self {
...
}
}
Additionally, Y
must implement the following (regardless of whether X
implements Clone
):
impl test_fuzz::Into<X> for Y {
fn into(self) -> X {
...
}
}
The definition of test_fuzz::Into
is identical to that of std::convert::Into
. The reason for using a non-standard trait is to avoid conflicts that could arise from blanket implementations of standard traits.
Generate corpus files when not running tests, provided the environment variable TEST_FUZZ_WRITE
is set. The default is to generate corpus files only when running tests, regardless of whether TEST_FUZZ_WRITE
is set. When running a target from outside its package directory, set TEST_FUZZ_MANIFEST_PATH
to the path of the package's Cargo.toml
file.
WARNING: Setting enable_in_production
could introduce a denial-of-service vector. For example, setting this option for a function that is called many times with different arguments could fill up the disk. The check of TEST_FUZZ_WRITE
is meant to provide some defense against this possibility. Nonetheless, consider this option carefully before using it.
Rather than call the target directly:
- construct a closure of type
FnOnce() -> R
, whereR
is the target's return type, so that calling the closure calls the target; - call
function
with the closure.
Calling the target in this way allows function
to set up the call's environment. This can be useful, e.g., for fuzzing Substrate externalities.
Do not try to auto-generate corpus files for the target.
Record the target's generic args when running tests, but do not generate corpus files and do not implement a fuzzing harness. This can be useful when the target is a generic function, but it is unclear what type parameters should be used for fuzzing.
The intended workflow is: enable only_generic_args
, then run cargo test
followed by cargo test-fuzz --display generic-args
. One of the resulting generic args might be usable as generic_args
's parameters
. Similarly, generic args resulting from cargo test-fuzz --display impl-generic-args
might be usable as impl_generic_args
's parameters
.
Note, however, that just because a target was called with certain parameters during tests, it does not imply the target's arguments are serializable/deserializable when those parameters are used. The results of --display generic-args
/--display impl-generic-args
are merely suggestive.
Treat the target as though its name is name
when adding a module to the enclosing scope. Expansion of the test_fuzz
macro adds a module definition to the enclosing scope. By default, the module is named as follows:
- If the target does not appear in an
impl
block, the module is namedtarget_fuzz
, wheretarget
is the name of the target. - If the target appears in an
impl
block, the module is namedpath_target_fuzz
, wherepath
is the path of theimpl
'sSelf
type converted to snake case and joined with_
.
However, use of this option causes the module to instead be named name_fuzz
. Example:
#[test_fuzz(rename = "bar")]
fn foo() {}
// Without the use of `rename`, a name collision and compile error would result.
mod foo_fuzz {}
The test_fuzz
macro allows Serde field attributes to be applied to function arguments. This provides another tool for dealing with difficult types.
The following is an example. Traits serde::Serialize
and serde::Deserialize
cannot be derived for Context
because it contains a Mutex
. However, Context
implements Default
. So applying #[serde(skip)]
to the Context
argument causes it to be skipped when serializing, and to take its default value when deserializing.
use std::sync::Mutex;
// Traits `serde::Serialize` and `serde::Deserialize` cannot be derived for `Context` because it
// contains a `Mutex`.
#[derive(Default)]
struct Context {
lock: Mutex<()>,
}
impl Clone for Context {
fn clone(&self) -> Self {
Self {
lock: Mutex::new(()),
}
}
}
#[test_fuzz::test_fuzz]
fn target(#[serde(skip)] context: Context, x: i32) {
assert!(x >= 0);
}
Note that when Serde field attributes are applied to an argument, the test_fuzz
macro performs no other conversions on the argument.
Whenever the test_fuzz
macro is used in an impl
block,
the impl
must be preceded with the test_fuzz_impl
macro. Example:
#[test_fuzz_impl]
impl Foo {
#[test_fuzz]
fn bar(&self, x: &str) {
...
}
}
The reason for this requirement is as follows. Expansion of the test_fuzz
macro adds a module definition to the enclosing scope. However, a module definition cannot appear inside an impl
block. Preceding the impl
with the test_fuzz_impl
macro causes the module to be added outside the impl
block.
If you see an error like the following, it likely means a use of the test_fuzz_impl
macro is missing:
error: module is not supported in `trait`s or `impl`s
test_fuzz_impl
currently has no options.
The cargo test-fuzz
command is used to interact with fuzz targets, and to manipulate their corpora, crashes, hangs, and work queues. Example invocations include:
-
List fuzz targets
cargo test-fuzz --list
-
Display target
foo
's corpuscargo test-fuzz foo --display corpus
-
Fuzz target
foo
cargo test-fuzz foo
-
Replay crashes found for target
foo
cargo test-fuzz foo --replay crashes
Usage: cargo test-fuzz [OPTIONS] [TARGETNAME] [-- <ARGS>...]
Arguments:
[TARGETNAME] String that fuzz target's name must contain
[ARGS]... Arguments for the fuzzer
Options:
--backtrace Display backtraces
--consolidate Move one target's crashes, hangs, and work queue to its corpus; to
consolidate all targets, use --consolidate-all
--cpus <N> Fuzz using at most <N> cpus; default is all but one
--display <OBJECT> Display corpus, crashes, generic args, `impl` generic args, hangs,
or work queue. By default, an uninstrumented fuzz target is used.
To display with instrumentation, append `-instrumented` to
<OBJECT>, e.g., --display corpus-instrumented.
--exact Target name is an exact name rather than a substring
--exit-code Exit with 0 if the time limit was reached, 1 for other
programmatic aborts, and 2 if an error occurred; implies --no-ui,
does not imply --run-until-crash or --max-total-time <SECONDS>
--features <FEATURES> Space or comma separated list of features to activate
--list List fuzz targets
--manifest-path <PATH> Path to Cargo.toml
--max-total-time <SECONDS> Fuzz at most <SECONDS> of time (equivalent to -- -V <SECONDS>)
--no-default-features Do not activate the `default` feature
--no-run Compile, but don't fuzz
--no-ui Disable user interface
-p, --package <PACKAGE> Package containing fuzz target
--persistent Enable persistent mode fuzzing
--pretty Pretty-print debug output when displaying/replaying
--replay <OBJECT> Replay corpus, crashes, hangs, or work queue. By default, an
uninstrumented fuzz target is used. To replay with instrumentation
append `-instrumented` to <OBJECT>, e.g., --replay
corpus-instrumented.
--reset Clear fuzzing data for one target, but leave corpus intact; to
reset all targets, use --reset-all
--resume Resume target's last fuzzing session
--run-until-crash Stop fuzzing once a crash is found
--slice <SECONDS> If there are not sufficiently many cpus to fuzz all targets
simultaneously, fuzz them in intervals of <SECONDS> [default:
1200]
--test <NAME> Integration test containing fuzz target
--timeout <TIMEOUT> Number of seconds to consider a hang when fuzzing or replaying
(equivalent to -- -t <TIMEOUT * 1000> when fuzzing)
--verbose Show build output when displaying/replaying
-h, --help Print help
-V, --version Print version
Try `cargo afl fuzz --help` to see additional fuzzer options.
Warning: These utilties are excluded from semantic versioning and may be removed in future versions of test-fuzz
.
The dont_care!
macro can be used to implement serde::Serialize
/serde::Deserialize
for types that are easy to construct and whose values you do not care to record. Intuitively, dont_care!($ty, $expr)
says:
- Skip values of type
$ty
when serializing. - Initialize values of type
$ty
with$expr
when deserializing.
More specifically, dont_care!($ty, $expr)
expands to the following:
impl serde::Serialize for $ty {
fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
where
S: serde::Serializer,
{
().serialize(serializer)
}
}
impl<'de> serde::Deserialize<'de> for $ty {
fn deserialize<D>(deserializer: D) -> std::result::Result<Self, D::Error>
where
D: serde::Deserializer<'de>,
{
<()>::deserialize(deserializer).map(|_| $expr)
}
}
If $ty
is a unit struct, then $expr
can be omitted. That is, dont_care!($ty)
is equivalent to dont_care!($ty, $ty)
.
The leak!
macro can help to serialize target arguments that are references and whose types implement the ToOwned
trait. It is meant to be used with the convert
option.
Specifically, an invocation of the following form declares a type LeakedX
, and implements the From
and test_fuzz::Into
traits for it:
leak!(X, LeakedX);
One can then use LeakedX
with the convert
option as follows:
#[test_fuzz::test_fuzz(convert = "&X, LeakedX")
An example where X
is Path
appears in conversion.rs in this repository.
More generally, an invocation of the form leak!($ty, $ident)
expands to the following:
#[derive(Clone, std::fmt::Debug, serde::Deserialize, serde::Serialize)]
struct $ident(<$ty as ToOwned>::Owned);
impl From<&$ty> for $ident {
fn from(ty: &$ty) -> Self {
Self(ty.to_owned())
}
}
impl test_fuzz::Into<&$ty> for $ident {
fn into(self) -> &'static $ty {
Box::leak(Box::new(self.0))
}
}
serialize_ref
and deserialize_ref
function similar to leak!
, but they are meant to be used wth Serde's serialize_with
and deserialize_with
field attributes (respectively).
fn serialize_ref<S, T>(x: &&T, serializer: S) -> Result<S::Ok, S::Error>
where
S: serde::Serializer,
T: serde::Serialize,
{
<T as serde::Serialize>::serialize(*x, serializer)
}
fn deserialize_ref<'de, D, T>(deserializer: D) -> Result<&'static T, D::Error>
where
D: serde::Deserializer<'de>,
T: serde::de::DeserializeOwned + std::fmt::Debug,
{
let x = <T as serde::de::Deserialize>::deserialize(deserializer)?;
Ok(Box::leak(Box::new(x)))
}
serialize_ref_mut
and deserialize_ref_mut
are similar to serialize_ref
and deserialize_ref
(respectively), expect they operate on mutable references instead of immutable ones.
The features in this section apply to the test-fuzz
package as a whole. Enable them in test-fuzz
's dependency specification as described in the The Cargo Book. For example, to enable the cast_checks
feature, use:
test-fuzz = { version = "*", features = ["cast_checks"] }
The test-fuzz
package currently supports the following features:
Use cast_checks
to automatically check target functions for invalid casts.
Note that this feature enables cast_checks
only for functions annotated with the test_fuzz
macro, not for the functions they call.
test-fuzz
can serialize target arguments in multiple Serde formats. The following are the features used to select a format.
cargo-test-fuzz
can auto-generate values for types that implement certain traits. If all of a target's argument types implement such traits, cargo-test-fuzz
can auto-generate corpus files for the target.
The traits that cargo-test-fuzz
currently supports and the values generated for them are as follows:
Trait(s) | Value(s) |
---|---|
Bounded |
T::min_value() , T::max_value() |
Bounded + Add + One |
T::min_value() + T::one() |
Bounded + Add + Div + Two |
T::min_value() / T::two() + T::max_value() / T::two() |
Bounded + Add + Div + Two + One |
T::min_value() / T::two() + T::max_value() / T::two() + T::one() |
Bounded + Sub + One |
T::max_value() - T::one() |
Default |
T::default() |
Key
Add
-core::ops::Add
Bounded
-num_traits::bounds::Bounded
Default
-std::default::Default
Div
-core::ops::Div
One
-num_traits::One
Sub
-core::ops::Sub
Two
-test_fuzz::runtime::traits::Two
(essentiallyAdd + One
)
During macro expansion:
- If
TEST_FUZZ_LOG
is set to1
, write all instrumented fuzz targets and module definitions to standard output. - If
TEST_FUZZ_LOG
is set to a crate name, write that crate's instrumented fuzz targets and module definitions to standard output.
This can be useful for debugging.
When running a target from outside its package directory, find the package's Cargo.toml
file at this location. One may need to set this environment variable when enable_in_production
is used.
Generate corpus files when not running tests for those targets for which enable_in_production
is set.
A target's arguments must implement the Clone
trait. The reason for this requirement is that the arguments are needed in two places: in a test-fuzz
-internal function that writes corpus files, and in the body of the target function. To resolve this conflict, the arguments are cloned before being passed to the former.
In general, a target's arguments must implement the serde::Serialize
and serde::Deserialize
traits, e.g., by deriving them. We say "in general" because test-fuzz
knows how to handle certain special cases that wouldn't normally be serializable/deserializable. For example, an argument of type &str
is converted to String
when serializing, and back to a &str
when deserializing. See also generic_args
and impl_generic_args
above.
The fuzzing harnesses that test-fuzz
implements do not initialize global variables. While execute_with
provides some remedy, it is not a complete solution. In general, fuzzing a function that relies on global variables requires ad-hoc methods.
These options are incompatible in the following sense. If a fuzz target's argument type is a type parameter, convert
will try to match the type parameter, not the type to which the parameter is set. Supporting the latter would seem to require simulating type substitution as the compiler would perform it. However, this is not currently implemented.
-
#[cfg(test)]
is not enabled for integration tests. If your target is tested only by integration tests, then consider usingenable_in_production
andTEST_FUZZ_WRITE
to generate a corpus. (Note the warning accompanyingenable_in_production
, however.) -
If you know the package in which your target resides, passing
-p <package>
tocargo test
/cargo test-fuzz
can significantly reduce build times. Similarly, if you know your target is called from only one integration test, passing--test <name>
can reduce build times. -
Rust won't allow you to implement
serde::Serialize
for other repositories' types. But you may be able to patch other repositories to make their types serializable. Also,cargo-clone
can be useful for grabbing dependencies' repositories. -
Serde attributes can be helpful in implementing
serde::Serialize
/serde::Deserialize
for difficult types.
We reserve the right to change the format of corpora, crashes, hangs, and work queues, and to consider such changes non-breaking.
test-fuzz
is licensed and distributed under the AGPLv3 license with the Macros and Inline Functions Exception. In plain language, using the test_fuzz
macro, the test_fuzz_impl
macro, or test-fuzz
's convenience functions and macros in your software does not require it to be covered by the AGPLv3 license.