Feature #20448: Make coverage event hooking C API public

• Status: Open
• Assignee: -
• Target version: -

[ms-tob (Matt S)]

Abstract

Gathering code coverage information is a well-known goal within software engineering. It is most commonly used to assess code coverage during automated testing. A lesser known use-case is coverage-guided fuzz testing, which will be the primary use-case presented in this issue. This issue exists to request that Ruby coverage event hooking be made part of its official, public C API.

Background

Ruby currently provides a number of avenues for hooking events or gathering coverage information:

1. The Coverage module
2. The TracePoint module
3. The rb_add_event_hook extension function

Unfortunately, none of these pieces of functionality solve this issue's specific use-case. The Coverage module is not a great fit for real-time coverage analysis with an unknown start and stop point. Coverage-guided fuzz testing requires this. The TracePoint module and rb_add_event_hook are not able to hook branch and line coverage events. Coverage-guided fuzz testing typically tracks branch events.

Proposal

The ultimate goal is to enable Ruby C extensions to process coverage events in real-time. I did some cursory investigation into the Ruby C internals to determine what it would take to achieve this, but I'm by no means an expert, so my list may be incomplete.

The good news is that much of this functionality already exists, but it's part of the private, internal-only C API.

1. Make RUBY_EVENT_COVERAGE_LINE and RUBY_EVENT_COVERAGE_BRANCH public: https://github.com/ruby/ruby/blob/v3_3_0/vm_core.h#L2182-L2184
   a. This would be an addition to the current public event types: https://github.com/ruby/ruby/blob/v3_3_0/include/ruby/internal/event.h#L32-L46
2. Allow initializing global coverage state so that coverage tracking can be fully enabled
   a. Currently, if Coverage.setup or Coverage.start is not called, then coverage events cannot be hooked. I do not fully understand why this is, but I believe it has something to do with rb_get_coverages and rb_set_coverages. If calls to rb_get_coverages return NULL (https://github.com/ruby/ruby/blob/v3_3_0/iseq.c#L641-L647, https://github.com/ruby/ruby/blob/v3_3_0/iseq.c#L864-L868), then coverage hooking will not be enabled. I believe the Coverage module initializes that state via a rb_set_coverages call here: https://github.com/ruby/ruby/blob/v3_3_0/ext/coverage/coverage.c#L112-L120.
   b. So, to achieve this goal, a C extension would need to be able to call rb_set_coverages or somehow initialize the global coverage state.

I've actually been able to achieve this functionality by calling undocumented features and defining RUBY_EVENT_COVERAGE_BRANCH:

#include <ruby.h>
#include <ruby/debug.h>

#define RUBY_EVENT_COVERAGE_BRANCH 0x020000

// ...

rb_event_flag_t events = RUBY_EVENT_COVERAGE_BRANCH;
rb_event_hook_flag_t flags = (
    RUBY_EVENT_HOOK_FLAG_SAFE | RUBY_EVENT_HOOK_FLAG_RAW_ARG
);
rb_add_event_hook2(
    (rb_event_hook_func_t) event_hook_branch,
    events,
    counter_hash,
    flags
);

If I call Coverage.setup(branches: true), and add this event hook, then branch hooking works as expected. rb_add_event_hook2 will still respect the RUBY_EVENT_COVERAGE_BRANCH value if its passed. But it would be better if I could rely on official functionality rather than undocumented features.

The above two points would be requirements for this functionality, but there's an additional nice-to-have:

3. Extend the public tracearg functionality to include additional coverage information
   a. Currently, tracearg offers information like rb_tracearg_lineno and rb_tracearg_path. It would be helpful if it also provided additional coverage information like coverage.c's column information and a unique identifier for each branch. Currently, I can only use (path, lineno) as a unique identifier for a branch because that's what's offered by the public API, but more information like column number would be helpful for uniquely identify branches. Since there can be multiple if statements on a single line, this can provide ambiguous identification for a branch event.

Use cases

This use-case was born out of a new coverage-guided Ruby fuzzer: https://github.com/trailofbits/ruzzy. You can read more about its implementation details here: https://blog.trailofbits.com/2024/03/29/introducing-ruzzy-a-coverage-guided-ruby-fuzzer/. You can also find the Ruby C extension code behind its implementation here: https://github.com/trailofbits/ruzzy/blob/v0.7.0/ext/cruzzy/cruzzy.c#L220-L231.

So, the primary use-case here is enabling real-time, coverage-guided fuzz testing of Ruby code. However, as mentioned in the abstract, gathering code coverage information is useful in many domains. For example, it could enable new workflows in standard unit/integration test coverage. It could also enable gathering coverage information in real-time as an application is running. I see this as the most generalized form of gathering code coverage information, and something like the Coverage module as a specialized implementation. Another example, https://bugs.ruby-lang.org/issues/20282 may be solved by this more generalized solution.

We are tracking this request downstream here: https://github.com/trailofbits/ruzzy/issues/9

Discussion

Fuzz testing is another tool in a testers toolbelt. It is an increasingly common way to improve software's robustness. Go has it built in to the standard library, Python has Atheris, Java has Jazzer, JavaScript has Jazzer.js, etc. OSS-Fuzz has helped identify and fix over 10,000 vulnerabilities and 36,000 bugs using fuzzing. Ruby deserves a good fuzzer, and improving coverage gathering would help achieve that goal.

The Coverage module, TracePoint module, and rb_add_event_hook function seem like they could fulfill this goal. However, after deeper investigation, none of them fit the exact requirements for this use-case.

See also

• https://bugs.ruby-lang.org/issues/20282
• https://github.com/google/atheris
  • https://security.googleblog.com/2020/12/how-atheris-python-fuzzer-works.html
• https://github.com/CodeIntelligenceTesting/jazzer/
  • https://www.code-intelligence.com/blog/java-fuzzing-with-jazzer
• https://go.dev/doc/security/fuzz/

[mame (Yusuke Endoh)]

First of all, it's great that you are creating a fuzzing tool for Ruby. Thank you for your work.

I understand that what you need is a hook for LINE and BRANCH events. Am I correct?

If so, did you try RUBY_EVENT_LINE? It is already a public event and should be almost equivalent to RUBY_EVENT_COVERAGE_LINE. One difference is that RUBY_EVENT_COVERAGE_LINE hook is only called for bytecodes built after Coverage.start is executed (i.e., after rb_get_coverages returns non-null), whereas RUBY_EVENT_LINE hook is called for all bytecodes including one built before Coverage.start. I don't know if this difference will be a problem for your use case, but if necessary, you could filter by file path.

RUBY_EVENT_COVERAGE_BRANCH is more complicated because it strongly depends on compiler implementation details. This feature requires an instrumented bytecode to be output at compile time, so it must be enabled before compilation of target Ruby code. Also, currently, column information is not left in the byte code. The current branch coverage measurement is achieved by separately bookkeeping the column information at compile time. I don't know if these can be cut out as a clean API.

[ms-tob (Matt S)]

I understand that what you need is a hook for LINE and BRANCH events. Am I correct?

I just need a hook for BRANCH events. You can find that C extension code here:

https://github.com/trailofbits/ruzzy/blob/v0.7.0/ext/cruzzy/cruzzy.c#L220-L231

And I manually define RUBY_EVENT_COVERAGE_BRANCH above with #define RUBY_EVENT_COVERAGE_BRANCH 0x020000.

RUBY_EVENT_COVERAGE_BRANCH is more complicated because it strongly depends on compiler implementation details. This feature requires an instrumented bytecode to be output at compile time, so it must be enabled before compilation of target Ruby code. Also, currently, column information is not left in the byte code. The current branch coverage measurement is achieved by separately bookkeeping the column information at compile time. I don't know if these can be cut out as a clean API.

I was able to figure out a workaround for this. By reviewing the Coverage module documentation I learned that you have to start coverage gathering, and then require a separate Ruby module. My C extension takes the following actions to perform this:

• Enable coverage gathering: https://github.com/trailofbits/ruzzy/blob/v0.7.0/ext/cruzzy/cruzzy.c#L195-L211
• Then, require the caller-provided module name: https://github.com/trailofbits/ruzzy/blob/v0.7.0/ext/cruzzy/cruzzy.c#L233

Due to this behavior, when fuzzing pure Ruby code (not C extensions), Ruzzy requires two separate scripts. A "tracer" script and a fuzzing harness: https://github.com/trailofbits/ruzzy#fuzzing-pure-ruby-code.

Having this behavior built into the language is extremely helpful. Without it, as you mentioned, we would have to manually instrument the bytecode at runtime to achieve coverage-guided fuzzing. This is what Atheris has to do to fuzz Python code, and it requires significantly more investment into the solution. So, I really appreciate that Ruby has this built in, even if its private, internal-only functionality.

[mame (Yusuke Endoh)]

I just need a hook for BRANCH events.

It's a hard mode ;-)

Ruzzy requires two separate scripts

I guess that the tracer script is always the same (except the file name of the test harness). If so, you may want to make it a library like ruzzy/tracer.rb. A user only need to write test_harness.rb, and can invoke with ruby -rruzzy/tracer test_harness.rb. This is just a matter of taste, though.

Now, I wonder what API design would be good. Can we make TracePoint support branch events? If so, I think coverage.so can be reimplemented on top of TracePoint.

How about an API like this one? (I am not sure if this is really implementable until someone implements it.)

TracePoint.enable_branch_tracepoints

TracePoint.new(:branch) do |tp|
  p tp.branch_id #=> Integer ID which is unique for each file path

  p tp.branch_from #=> [first_lineno, first_column, last_lineno, last_column]
  p tp.branch_to #=> [first_lineno, first_column, last_lineno, last_column]
end.enable

load "target.rb"

TracePoint.disable_branch_tracepoints

@ko1 (Koichi Sasada) What do you think?
@ms-tob (Matt S) Would it be sufficient for your use case?

[mame (Yusuke Endoh)]

Oh, this API does not allow to get information of branch_from and branch_to for branches that are never fired. It is insufficient to reimplement coverage.so on TracePoint. We have to actually create a patch and identify the appropriate API.

[ms-tob (Matt S)]

Sorry for the delay, I've been considering different APIs and reevaluating Ruzzy's coverage collection. I think I've identified that Ruzzy's current coverage collection mechanism is flawed, so perhaps we should not use that as the goal. I'll explain the goal at a high-level, show where I think Ruzzy's implementation is flawed, and I'd be curious to hear what you think.

For fuzzing coverage collection the ultimate goal will be to instrument basic blocks. To instrument basic blocks we need a unique identifier for each one. These identifiers do not have to be consistent across fuzzing runs, just unique within a single fuzzing run (i.e. an invocation of the fuzzing program). For Ruby, I think this would look something like the following:

def main
  # Basic block: BB1, ignore root block for now, it's a special case
  # ...
  if username == "John"
    # Basic block BB2
    # ...
  elsif username == "Jane"
    # Basic block BB3
    # ...
  else
    # Basic block BB4
    # ...
  end
  # ...
  unless username == "David"
    # Basic block BB5
    # ...
  end
  # ...
  case username
  when "Danielle"
    # Basic block BB6
    # ...
  when "Mark"
    # Basic block BB7
    # ...
  else
    # Basic block BB8
    # ...
  end
  # ...
end

Forgive me if I'm missing some branching constructs, but the goal here is to maximize branch coverage. So we'd like to be able to notify the fuzzer that it has generated an input that increased coverage. In other words, it has found a new branch. This is accomplished by identifying new basic blocks after a branch event. Atheris achieves this by rewriting the Python bytecode and inserting a function call with a unique identifier in the branch's target basic block. It does this for all conditional jumps, i.e. branches.

Currently, Ruzzy is only instrumenting (filepath, lineno) tuples during branch events (RUBY_EVENT_COVERAGE_BRANCH). I think this is effectively producing instrumentation like the following:

def main
  # ...
  if username == "John" # Basic block BB1
    # ...
  elsif username == "Jane"
    # ...
  else
    # ...
  end
  # ...
  unless username == "David" # Basic block BB2
    # ...
  end
  # ...
  case username # Basic block BB3 (are case statements included in branch events?)
  when "Danielle"
    # ...
  when "Mark"
    # ...
  else
    # ...
  end
  # ...
end

I'll have to do some testing to confirm if this is how RUBY_EVENT_COVERAGE_BRANCH works or not. Correct me if I'm wrong, but I do not think hooking RUBY_EVENT_COVERAGE_BRANCH will work because it does not provide information on which branch is being taken.

The branches information in the Coverage module is very close to what we need. It provides a unique identifier for each basic block. The only deficiency is that we need this information in realtime as a branch event happens.

The debug.h module publicly provides us with a rb_trace_arg_t during branch events, but we don't have access to any of the rb_trace_arg_struct internals. The Coverage.result functionality is ultimately provided with this internal coverage information by a call to rb_get_coverages. It then generates unique identifiers for basic blocks after coverage information has already been gathered.

I wonder, is there a way to generalize this functionality such that the Coverage module can still use it, and a hooking API is provided to publicly expose this coverage gathering in realtime? I think something like the following would work:

TracePoint.enable_branch_tracepoints

TracePoint.new(:branch) do |tp|
  p tp.branch_id #=> Integer ID which is unique for each file path
  p tp.branch_target #=> [id, first_lineno, first_column, last_lineno, last_column], Another unique ID for branch target
end.enable

load "target.rb"

TracePoint.disable_branch_tracepoints

This is relatively close to the existing Coverage branches result. What do you think?

[ms-tob (Matt S)]

TL;DR: I've been further researching how fuzzers gather coverage information, and what types of coverage are most useful for fuzzing. I think the best course of action regarding this request would be to implement a fairly general coverage instrumentation approach that can be used to implement different types of coverage.

Based on my research, I've found that there are four commonly used types of coverage:

1. Function
2. Basic block
3. Edge
4. Path

libFuzzer's SanitizerCoverage tool offers three coverage instrumentation points: func, bb, and edge. So these would correspond to (1), (2), and (3) above.

AFL's coverage instrumentation implements edge coverage. This corresponds to (3) above.

I have not seen any practical fuzzers that implement (4) above. I believe this is likely due to the state explosion of trying to track all unique paths through a program. So, if we can implement functionality in the TracePoint module to gather coverage information for instrumentation points (1), (2), and (3), then I think we will cover the vast majority of use-cases. libFuzzer and AFL are far and away the two most popular fuzzers. Instead of focusing solely on my specific use-case, I think it's a good idea to enable a generalized solution that others can test out different fuzzing strategies with.

The good news is I think supporting (1) above is already possible with the call, a_call, b_call, and c_call TracePoint events. However, function based instrumentation is the least useful in my opinion. I see function -> basic block -> edge tracking as least to most granular. The tradeoff here is higher granularity requires potentially more memory and compute time, although I think it's negligible in this case.

So, if Ruby already supports events for (1), and we're discarding (4), then that leaves us with (2) and (3). Fortunately, both (2) and (3) can be implemented based on a control-flow graph (CFG) generated from branch events. The main difference here is that basic block coverage is simply branch destinations, whereas edge coverage is branch sources + branch destinations. In other words, basic blocks are simply (branch_dst) and edges are (branch_src, branch_dst). This is described well in AFL's coverage measurement technical documentation.

I've been reading up on this quite a bit, so I've picked out a few quotations to back up my claims:

... tracking accurate path coverage is infeasible in practice, but tracking edge and block coverage are possible. The edge coverage provided by AFL provides more information than block coverage solutions.

• CollAFL: Path Sensitive Fuzzing

AFL [60] leverages edge coverage (a.k.a. branch coverage or transition coverage), and libFuzzer [47] supports both edge and block coverage.

• Not All Coverage Measurements Are Equal: Fuzzing by Coverage Accounting for Input Prioritization

The number of unique paths (#paths) continues to be a common performance measure for greybox fuzzers [17, 18, 68] despite its obvious flaws [36, 38]. The number of unique edges (#edges), reported as map size by AFL-based fuzzers, is often used as a proxy for branch coverage. AFL maintains a fixed-size hashmap containing an entry for every tuple of conditional jumps that are sequentially exercised in the program.

• On the Reliability of Coverage-Based Fuzzer Benchmarking

Branch coverage is a straightforward yet effective enhancement over block coverage, which is the most basic one that can only tell which code block is visited. By involving the code block preceding the currently visited one, branch coverage can differentiate the visits of the same code block from different predecessors. Branch here means an edge from one code block to another one. Ideally, branch coverage should be measured as a tuple (prev_block, cur_block), where prev_block and cur_block stand for the previous block ID and the current block ID, respectively. In practice, branch coverage is usually measured by hashing this tuple (as key) into a hash table (e.g., a hit_count map).

• Be Sensitive and Collaborative: Analyzing Impact of Coverage Metrics in Greybox Fuzzing

I believe we can achieve both basic block and edge coverage instrumentation with the following API:

TracePoint.enable_branch_tracepoints

TracePoint.new(:branch) do |tp|
  p tp.branch_src #=> [src_id, src_first_lineno, src_first_column, src_last_lineno, src_last_column]
  p tp.branch_dst #=> [dst_id, dst_first_lineno, dst_first_column, dst_last_lineno, dst_last_column]
end.enable

load "target.rb"

TracePoint.disable_branch_tracepoints

In this case, the src is the location in code of the branch statement (e.g. the if statement), and the dst is the location in code where execution will continue after the branch statement (i.e. which branch is taken). If a fuzzer would like to track basic blocks, they can use only branch_dst. If they'd like to track edges, they can use both branch_src and branch_dst. This would allow fuzzers flexibility in their implementation and allow them to experiment with different instrumentation points to optimize the fuzzing process. It would also allow Ruby fuzzers to be built on both AFL and libFuzzer, which enables further experimentation and options.

One small edge case: we will likely want to consider critical edges as well. This is an implementation detail, and shouldn't impact the overall API of the functionality.

I think this would be a good, generalized solution that many parties interested in fuzzing could experiment with. Let me know what you think of this proposal. I have a much better understanding of how fuzzers implement and use coverage information now, so I hope this additional information helps.

[mame (Yusuke Endoh)]

Thanks for the research. So, TracePoint.new(:branch) seems good enough for the use case of fuzzer, right?

But with just this API, there is no way to get the list of all edges, including ones that are not fired. This is needed for branch coverage because it need to warn branches that is not fired during test. Is that not a problem for fuzzer?

[ms-tob (Matt S)]

Thanks for the research. So, TracePoint.new(:branch) seems good enough for the use case of fuzzer, right?

Yes, TracePoint.new(:branch) will work. It would be nice to have a C interface as well, but it's not strictly necessary. The coverage information will need to make its way to a Ruby C extension anyway, so calling the interface from C would be slightly easier.

But with just this API, there is no way to get the list of all edges, including ones that are not fired. This is needed for branch coverage because it need to warn branches that is not fired during test. Is that not a problem for fuzzer?

Yes, that's correct. Our fuzzer implementation is only concerned with edges that have been executed, not necessarily all possible edges. Knowing all possible edges, including those which haven't been executed during a specific run, can be helpful for certain types of analyses, but not for our specific use-case.