fuzzer: excersizes query code on xml doc

[nathaniel-brough]

This change increases code coverage significantly during fuzzing as it excersizes the query code against an arbitrary file. I've also updated the fuzzer to use LLVM's FuzzedDataProvider which makes the fuzzer a little easier to understand/reason about.

[zeux]

It would be better if this change didn’t change pugixml.hpp; if the enumeration gets extended it will require changes in fuzzing code anyway so it seems better to explicitly switch/case or hardcode boolean as last enum member in the fuzzing harness.

Also the change should be converted to Allman brace style to fit the convention in the rest of the code.

[nathaniel-brough]

if the enumeration gets extended it will require changes in fuzzing code

I might be missing something, but it was my intention that this would not be required. As long as kMaxValue is aliased to the last element in the enum it shouldn't require updates to the fuzzer (see comments on ConsumeEnum).

I'm more than happy to refactor all of this slightly to not require changes to the main library header files. Although the alternatives are slightly less readable, and would definitely require updates to the fuzzer if the enum is extended.

[nathaniel-brough]

The kMaxValue (modification to pugixml.hpp) is required ConsumeEnum to work correctly.

[nathaniel-brough]

For posterity here is where kMaxValue is actually used by FuzzedDataProvider::ConsumeEnum https://github.com/llvm/llvm-project/blob/main/compiler-rt/include/fuzzer/FuzzedDataProvider.h#L298

[nathaniel-brough]

Also the change should be converted to Allman brace style to fit the convention in the rest of the code.

Not a problem, do you mind if I add a .clang-format file as a separate commit with the Allman style braces? This way most editors will autoformat to Allman style rather than some other default.

[zeux]

I might be missing something, but it was my intention that this would not be required. As long as kMaxValue is aliased to the last element in the enum it shouldn't require updates to the fuzzer (see comments on ConsumeEnum).

Sorry what I meant is that the fuzzing code right now has explicit calls to evaluate_X for each X, which is of course correct. If new types are added, this will expand the surface of the API which would require changing the fuzzing code anyway. So it seems preferable to avoid changing the header in this case and just hardcode the enum selection via an array lookup or something similar.

As an aside, just out of curiosity, is fuzzing efficiency better/worse/the same when using a specific fixed XML document? (eg some existing .xml file from the repository, or we could add a new one for this specific purpose)

I'm wondering this because it feels like a lot of generated documents will not really be valid or interesting XML, and as such a lot of queries would terminate early or not run at all, whereas if we knew the document was valid and had a variety of nodes with different types and structure, we would be able to directly fuzz the evaluation engine on known-interesting document and unknown query.

[nathaniel-brough]

Oh yeah I see what you are saying. I'll make some changes so that the header file isn't modified.

As an aside, just out of curiosity, is fuzzing efficiency better/worse/the same when using a specific fixed XML document? (eg some existing .xml file from the repository, or we could add a new one for this specific purpose)

The overall performance increase in code coverage >20%. But it's worth noting that this is using some other changes that I've made in the google/oss-fuzz configuration for this repo (I'll open a PR with oss-fuzz soon).

• Concatenate fuzz_parse.dict and fuzz_xpath.dict so that the XML and query components are in the dictionary.
• Merge the fuzz_parse and fuzz_xpath seed corpus

I haven't tested this change as standalone (i.e. without corpus/dictionary updates), but with the fairly minimal modifications (mentioned above), it's a fairly decent improvement.

But you raise a valid point that to really get the best performance a good set of handwritten seed corpus xml files would certainly help cover the most code for a query.

[nathaniel-brough]

Ok so this PR no longer changes the pugixml.hpp file and instead fuzzes this by casting a Consumed Integral back to the enum type.

You can find the PR with OSS-fuzz I mentioned earlier here google/oss-fuzz#9900.

I'll have a go at writing a few specific entries for the seed corpus later this week. But with the combination of these two PRs and fuzzing for a couple of hours on my laptop I've been able to increase code coverage by a bit over 20%.

[zeux]

Thanks!