Fix uregex_open_fuzzer: hardcoded flags and unaligned memory access#3919
Fix uregex_open_fuzzer: hardcoded flags and unaligned memory access#3919OwenSanzas wants to merge 2 commits intounicode-org:mainfrom
Conversation
…om fuzz data The harness reads UListFormatterType and UListFormatterWidth directly from fuzz data via memcpy into enum variables without validation. Loading arbitrary bit patterns into C++ enum types is undefined behavior, confirmed by UndefinedBehaviorSanitizer: runtime error: load of value 2125315823, which is not a valid value for type 'UListFormatterType' The fix reads into int32_t first, then clamps to valid enum ranges before casting. The redundant third createInstance call (which was the only one using validated values) is removed since the second call now uses validated values.
|
Hi @OwenSanzas thanks for your proposed changes -- but we are wondering... Why are you trying to "fix" the fuzzer harness code? It is not production code, so does not usually come under the same kind of scrutiny. There are also some things that we probably don't want to change because we would lose fuzzer test coverage of library code, but before doing real reviews we wanted to ask what you were trying to achieve here. |
|
Hi @markusicu, thanks for asking! We're a security research team working on fuzz harness quality. We've found that subtle harness issues can silently waste fuzzing resources or even mask real bugs — for example, the OpenSSL For this specific fuzzer, the We're definitely not trying to reduce coverage — flags=0 is still a possible input, so nothing is lost. Happy to adjust anything if you have concerns! |
|
Ok, we will take a look. |
|
Sorry about the template, I'll keep it next time. Thanks for taking a look! |
2 participants
Summary
Fix two issues in
uregex_open_fuzzer.cppthat limit fuzzing effectiveness.1. Hardcoded flags parameter
The fuzzer always passes
0for theflagsparameter touregex_open(), meaning it only ever tests regex compilation with default flags. The regex engine supportsUREGEX_CASE_INSENSITIVE,UREGEX_MULTILINE,UREGEX_DOTALL,UREGEX_COMMENTS, and other flags that exercise different code paths — none of which are being tested.Fix: Derive flags from the first byte of fuzzer input instead of hardcoding
0.2. Unaligned memory access
The fuzzer casts
datadirectly tochar16_t*viareinterpret_cast<const char16_t*>(data). Sincedatais not guaranteed to be 2-byte aligned, this causes undefined behavior and faults on architectures with strict alignment requirements (e.g., ARM).Fix: Copy the pattern data to a properly aligned buffer using
memcpy, consistent with the approach used in other ICU fuzzers.Coverage comparison (60 seconds, empty seed corpus, ASan, libFuzzer)
The +18% edge coverage increase confirms that the hardcoded flags were preventing exploration of significant portions of the regex engine.