Note: Writing grammar fuzzers with libprotobuf-mutator requires greater effort than writing fuzzers with libFuzzer alone. If you run into problems, send an email to fuzzing@chromium.org for help.
Prerequisites: Knowledge of libFuzzer in Chromium and basic understanding of Protocol Buffers.
This document will walk you through:
libprotobuf-mutator is a package that allows libFuzzer’s mutation engine to manipulate protobufs. This allows libFuzzer's mutations to be more specific to the format it is fuzzing and less arbitrary. Below are some good use cases for libprotobuf-mutator:
testing/libfuzzer/fuzzers/url_parse_proto_fuzzer.cc, and testing/libfuzzer/fuzzers/url.proto. Its build configuration can be found in testing/libfuzzer/fuzzers/BUILD.gn. We also provide a walkthrough on how to do this in the section after the next.In the next section, we discuss building a fuzzer that targets code that accepts an already existing protobuf definition. In the section after that, we discuss how to write and build grammar-based fuzzers using libprotobuf-mutator. Interested readers may also want to look at this example of a libprotobuf-mutator fuzzer that is even more trivial than url_parse_proto_fuzzer.
This is almost as easy as writing a standard libFuzzer-based fuzzer. You can look at override_lite_runtime_plugin_test_fuzzer for an example of a working example of this (don't copy the line adding “//testing/libfuzzer:no_clusterfuzz” to additional_configs). Or you can follow this walkthrough:
Start by creating a fuzz target. This is what the .cc file will look like:
// my_fuzzer.cc #include "testing/libfuzzer/proto/lpm_interface.h" // Assuming the .proto file is path/to/your/proto_file/my_proto.proto. #include "path/to/your/proto_file/my_proto.pb.h" DEFINE_PROTO_FUZZER( const my_proto::MyProtoMessage& my_proto_message) { targeted_function(my_proto_message); }
The BUILD.gn definition for this target will be very similar to regular libFuzzer-based fuzzer_test. However it will also have libprotobuf-mutator in its deps. This is an example of what it will look like:
// You must wrap the target in "use_libfuzzer" since trying to compile the // target without use_libfuzzer will fail (for reasons alluded to in the next // step), which the commit queue will try. if (use_libfuzzer) { fuzzer_test("my_fuzzer") { sources = [ "my_fuzzer.cc" ] deps = [ // The proto library defining the message accepted by // DEFINE_PROTO_FUZZER(). ":my_proto", "//third_party/libprotobuf-mutator", ... ] } }
There‘s one more step however. Because Chromium doesn’t want to ship to users the full protobuf library, all .proto files in Chromium that are used in production contain this line: option optimize_for = LITE_RUNTIME But this line is incompatible with libprotobuf-mutator. Thus, we need to modify the proto_library build target so that builds when fuzzing are compatible with libprotobuf-mutator. To do this, change your proto_library to fuzzable_proto_library (don't worry, this works just like proto_library when use_libfuzzer is false) like so:
import("//third_party/libprotobuf-mutator/fuzzable_proto_library.gni") fuzzable_proto_library("my_proto") { ... }
And with that we have completed writing a libprotobuf-mutator fuzz target for Chromium code that accepts protobufs.
Once you have in mind the code you want to fuzz and the format it accepts, you are ready to start writing a libprotobuf-mutator fuzzer. Writing the fuzzer will have three steps:
LITE_RUNTIME).Create a new .proto using proto2 or proto3 syntax and define a message that you want libFuzzer to mutate.
syntax = "proto2"; package my_fuzzer; message MyProtoFormat { // Define a format for libFuzzer to mutate here. }
See testing/libfuzzer/fuzzers/url.proto for an example of this in practice. That example has extensive comments on URL syntax and how that influenced the definition of the Url message.
Create a new .cc and write a DEFINE_PROTO_FUZZER function:
// Needed since we use getenv(). #include <stdlib.h> // Needed since we use std::cout. #include <iostream> #include "testing/libfuzzer/proto/lpm_interface.h" // Assuming the .proto file is path/to/your/proto_file/my_format.proto. #include "path/to/your/proto_file/my_format.pb.h" // Put your conversion code here (if needed) and then pass the result to // your fuzzing code (or just pass "my_format", if your target accepts // protobufs). DEFINE_PROTO_FUZZER(const my_fuzzer::MyFormat& my_proto_format) { // Convert your protobuf to whatever format your targeted code accepts // if it doesn't accept protobufs. std::string native_input = convert_to_native_input(my_proto_format); // You should provide a way to easily retreive the native input for // a given protobuf input. This is useful for debugging and for seeing // the inputs that cause targeted_function to crash (which is the reason we // are here!). Note how this is done before targeted_function is called // since we can't print after the program has crashed. if (getenv("LPM_DUMP_NATIVE_INPUT")) std::cout << native_input << std::endl; // Now test your targeted code using the converted protobuf input. targeted_function(native_input); }
This is very similar to the same step in writing a standard libFuzzer fuzzer. The only real differences are accepting protobufs rather than raw data and converting them to the desired format. Conversion code can‘t really be explored in this guide since it is format-specific. However, a good example of conversion code (and a fuzz target) can be found in testing/libfuzzer/fuzzers/url_parse_proto_fuzzer.cc. That example thoroughly documents how it converts the Url protobuf message into a real URL string. A good convention is printing the native input when the LPM_DUMP_NATIVE_INPUT env variable is set. This will make it easy to retreive the actual input that causes the code to crash instead of the protobuf version of it (eg you can get the URL string that causes an input to crash rather than a protobuf). Since it is only a convention it is strongly recommended even though it isn’t necessary. You don't need to do this if the native input of targeted_function is protobufs. Beware that printing a newline can make the output invalid for some formats. In this case you should use fflush(0) since otherwise the program may crash before native_input is actually printed.
Define a fuzzer_test target and include your protobuf definition and libprotobuf-mutator as dependencies.
import("//testing/libfuzzer/fuzzer_test.gni") import("//third_party/protobuf/proto_library.gni") fuzzer_test("my_fuzzer") { sources = [ "my_fuzzer.cc" ] deps = [ ":my_format_proto", "//third_party/libprotobuf-mutator" ... ] } proto_library("my_format_proto") { sources = [ "my_format.proto" ] }
See testing/libfuzzer/fuzzers/BUILD.gn for an example of this in practice.
If you have messages that are defined recursively (eg: message Foo has a field of type Foo), make sure to bound recursive calls to code converting your message into native input. Otherwise you will (probably) end up with an out of memory error. The code coverage benefits of allowing unlimited recursion in a message are probably fairly low for most targets anyway.
Remember that proto definitions can be changed in ways that are backwards compatible (such as adding explicit values to an enum). This means that you can make changes to your definitions while preserving the usefulness of your corpus. In general adding fields will be backwards compatible but removing them (particulary if they are required) is not.
Make sure you understand the meaning of the different protobuf modifiers such as oneof and repeated as they can be counter-intuitive. oneof means “At most one of” while repeated means “At least zero”. You can hack around these meanings if you need “at least one of” or “exactly one of” something. For example, this is the proto code for exactly one of: MessageA or MessageB or MessageC:
message MyFormat { oneof a_or_b { MessageA message_a = 1; MessageB message_b = 2; } required MessageC message_c = 3; }
And here is the C++ code that converts it.
std::string Convert(const MyFormat& my_format) { if (my_format.has_message_a()) return ConvertMessageA(my_format.message_a()); else if (my_format.has_message_b()) return ConvertMessageB(my_format.message_b()); else // Fall through to the default case, message_c. return ConvertMessageC(my_format.message_c()); }
LPM makes it straightforward to write a fuzzer for code that needs multiple inputs. The steps for doing this are similar to those of writing a grammar based fuzzer, except in this case the grammar is very simple. Thus instructions for this use case are given below. Start by creating the proto file which will define the inputs you want:
// my_fuzzer_input.proto syntax = "proto2"; package my_fuzzer; message FuzzerInput { required bool arg1 = 1; required string arg2 = 2; optional int arg3 = 1; }
In this example, the function we are fuzzing requires a bool and a string and takes an int as an optional argument. Let's define our fuzzer harness:
// my_fuzzer.cc #include "testing/libfuzzer/proto/lpm_interface.h" // Assuming the .proto file is path/to/your/proto_file/my_fuzzer_input.proto. #include "path/to/your/proto_file/my_proto.pb.h" DEFINE_PROTO_FUZZER( const my_proto::FuzzerInput& fuzzer_input) { if (fuzzer_input.has_arg3()) targeted_function_1(fuzzer_input.arg1(), fuzzer_input.arg2(), fuzzer_input.arg3()); else targeted_function_2(fuzzer_input.arg1(), fuzzer_input.arg2()); }
Then you must define build targets for your fuzzer harness and proto format in GN, like so:
import("//testing/libfuzzer/fuzzer_test.gni") import("//third_party/protobuf/proto_library.gni") fuzzer_test("my_fuzzer") { sources = [ "my_fuzzer.cc" ] deps = [ ":my_fuzzer_input", "//third_party/libprotobuf-mutator" ... ] } proto_library("my_fuzzer_input") { sources = [ "my_fuzzer_input.proto" ] }
Protobuf has a field rule repeated that is useful when a fuzzer needs to accept a non-fixed number of inputs (see mojo_parse_messages_proto_fuzzer, which accepts an unbounded number of mojo messages as an example). Protobuf version 2 also has optional and required field rules that some may find useful.
Once you have written a fuzzer with libprotobuf-mutator, building and running it is pretty much the same as if the fuzzer were a standard libFuzzer-based fuzzer (with minor exceptions, like your seed corpus must be in protobuf format).
Check out some of the existing proto fuzzers. Not only will they be helpful examples, it is possible that format you want to fuzz is already defined or partially defined by an existing proto definition (if you are writing a grammar fuzzer).
DEFINE_BINARY_PROTO_FUZZER can be used instead of DEFINE_PROTO_FUZZER (or DEFINE_TEXT_PROTO_FUZZER) to use protobuf‘s binary format for the corpus. This will make it hard/impossible to modify the corpus manually (i.e. when not fuzzing). However, protobuf’s text format (and by extension DEFINE_PROTO_FUZZER) is believed by some to come with a performance penalty compared to the binary format. We've never seen a case where this penalty was important, but if profiling reveals that protobuf deserialization is the bottleneck in your fuzzer, you may want to consider using the binary format. This will probably not be the case.