Abstract: Of coverage-guided fuzzing’s three main components: (1) testcase generation, (2) code coverage tracing, and (3) crash triage, code coverage tracing is a dominant source of overhead. Coverage-guided fuzzers trace every testcase’s code coverage through either static or dynamic binary instrumentation, or more recently, using hardware support. Unfortunately, tracing all testcases incurs significant performance penalties—even when the overwhelming majority of testcases and their coverage information are discarded because they do not increase code coverage. To eliminate needless tracing by coverage-guided fuzzers, we introduce the notion of coverage-guided tracing. Coverage-guided tracing leverages two observations: (1) only a fraction of generated testcases increase coverage, and thus require tracing; and (2) coverage-increasing testcases become less frequent over time. Coverage-guided tracing works by encoding the current frontier of code coverage in the target binary so that it self-reports when a testcase produces new coverage—without tracing. This acts as a filter for tracing; restricting the expense of tracing to only coverage-increasing testcases. Thus, coverage-guided tracing chooses to tradeoff increased coverage-increasing-testcase handling time for the ability to execute testcases initially at native speed. To show the potential of coverage-guided tracing, we create an implementation based on the static binary instrumentor Dyninst called UnTracer. We evaluate UnTracer using eight real-world binaries commonly used by the fuzzing community. Experiments show that after only an hour of fuzzing, UnTracer’s average overhead is below 1%, and after 24-hours of fuzzing, UnTracer approaches 0% overhead, while tracing every testcase with popular white- and black-box-binary tracers AFL-Clang, AFL-QEMU, and AFL-Dyninst incurs overheads of 36%, 612%, and 518%, respectively.
The basic idea of UnTracer is to patch the binary so that the portions that are covered are replaced by uninstrumented (and hence faster) code.
Edit: apparently, there is this old paper at PLDI 01 that talks about the same thing – A Framework for Reducing the Cost of Instrumented Code.