diff --git a/doc/note/fuzzing.md b/doc/note/fuzzing.md new file mode 100644 index 00000000..3342f01e --- /dev/null +++ b/doc/note/fuzzing.md @@ -0,0 +1,34 @@ +# Fuzzing + +Wuffs has been fuzzed (on [OSS-Fuzz](https://google.github.io/oss-fuzz/)) since +[February +2018](https://github.com/google/oss-fuzz/commit/27f374480b1e7c780d161b7a6727c23cc03419c5). + +As of November 2019, this has not yet found any memory-safety bugs in Wuffs. +Compile-time safety checks seem to work. + +However, it has found a few correctness bugs, based on generating random input. +Wuffs' fuzzer programs will [intentionally +segfault](https://github.com/google/wuffs/blob/7ec252876541ec203659949450fafddc148b606e/fuzz/c/fuzzlib/fuzzlib.c#L61) +if the Wuffs library returns an "internal" error message. + +For example, Wuffs [won't let you write `x -= +1`](/doc/note/interval-arithmetic.md#i--i--1-doesnt-compile) unless you can +prove that it won't underflow. The programmer might 'know' that `x` is always +positive at some point in their program, but to satisfy Wuffs' [bounds +checker](/doc/note/bounds-checking.md), they have to explicitly write something +like: + +``` +if x > 0 { + x -= 1 +} else { + // Unreachable: x is always positive because blah blah blah. + return "#internal error: etc" +} +``` + +Hitting the else branch means that there's a bug: what the programmer 'knew' +was incorrect. These aren't security bugs per se: decoding an image would +produce the wrong pixels, or abort early, instead of leading to RCE (Remote +Code Execution). But fuzzing Wuffs has still been useful.