Add MutatedTransform to the input type in TMinMutationalStage (#1251)

[am009]

No description provided.

[am009]

I'm not very sure about modifications (three lines of code) in the second commit (12b3849). The TMinMutationalStage will construct a new Testcase with the minimized input and replace the original one. I guess the post-operation is needed, for example, to add metadata to the new testcase.

[tokatoka]

shouldn't this be input_transformed.len()?

[am009]

This is one of the awkward places.

The input_transformed is now of the wrapped type MutatedTransform<Input, State>. I think here we want the length of the internal input (if I understand correctly). If we use try_transform_into to get the internal input (so that we can call .len()), MutatedTransform<Input, State> will be consumed. But we still need to keep MutatedTransform<Input, State> to pass into the mutator.

[tokatoka]

ok got it

[tokatoka]

I'm not very sure about modifications (three lines of code) in the second commit (12b3849). The TMinMutationalStage will construct a new Testcase with the minimized input and replace the original one. I guess the post-operation is needed, for example, to add metadata to the new testcase.

i think

            post.post_exec(state, corpus_idx)?;

already takes care of that post-operation right?

[am009]

I'm not very sure about modifications (three lines of code) in the second commit (12b3849). The TMinMutationalStage will construct a new Testcase with the minimized input and replace the original one. I guess the post-operation is needed, for example, to add metadata to the new testcase.

i think

already takes care of that post-operation right?

I think post.post_exec(state, corpus_idx)?; handles the corpus_id returned by fuzzer.process_execution. Is that corpus_id equal to the base_corpus_id?

I'm not sure about the behavior of fuzzer.process_execution. Will TMinMutationalStage inserts a new testcase into the corpus (and returns the new corpus_id) if the input is interesting like normal MutationalStage? This behavior does not seem useful for the minimization process.

[tokatoka]

I think post.post_exec(state, corpus_idx)?; handles the corpus_id returned by fuzzer.process_execution. Is that corpus_id equal to the base_corpus_id?

No
Base could be

1. The initial testcase that you start minimizing with
   for this one you don't have to call post_exec() right?
   Or,
2. The minimized testcase that you shrinked from 1)
   for this one you already called post_exec at line 162

Will TMinMutationalStage inserts a new testcase into the corpus (and returns the new corpus_id)

It's not really insert it's replace

[tokatoka]

wait I got it.

i think you should replace post.post_exec() with base_post.post_exec().

[am009]

No Base could be

1. The initial testcase that you start minimizing with
   for this one you don't have to call post_exec() right?
   Or,
2. The minimized testcase that you shrinked from 1)
   for this one you already called post_exec at line 162

But the base is of type Input, not Testcase<Input>. And there is a new test case constructed before the replacement. Metadata is probably lost? (and the post-operation are used to maintain the metadata, at least for StringIdentificationMetadata.)

It's not really insert it's replace

Yeah, the replacement is happening out of the loop, but there is also a corpus_idx returned by the fuzzer.process_execution inside the loop. Is that a new index for a new testcase?

[tokatoka]

Now I understand

Metadata is probably lost? (and the post-operation are used to maintain the metadata, at least for StringIdentificationMetadata.)

Yes you are right

[tokatoka]

Yeah, the replacement is happening out of the loop, but there is also a corpus_idx returned by the fuzzer.process_execution inside the loop. Is that a new index for a new testcase?

I think It's the same in index as before.

[tokatoka]

can you try not to use unwrap() here?

[am009]

OK, I have updated.
Now it will return an error when the option is empty:

            // perform the post operation for the new testcase, e.g. to update metadata.
            // base_post should be updated along with the base (and is no longer None)
            base_post
                .ok_or_else(|| Error::empty_optional("Failed to get the MutatedTransformPost"))?
                .post_exec(state, Some(base_corpus_idx))?;

[tokatoka]

is it possible to avoid this clone() ?

[am009]

The function MutatedTransformPost.post_exec consumes self. For example, the post is a metadata, and its ownership is transferred to the metadata map in the post_exec. So I guess it can't be avoided?

[domenukk]

I think you could set a bool to true here and then do base_post conversion and post_exec below, after calling post.post_exec

[am009]

So do we still need to clone? I think the post object is no longer valid after post_exec, because it is comsumed.

[domenukk]

ooh...
Well I guess in that case not much we can do..

[tokatoka]

no i removed it

[tokatoka]

@am009
I removed it
can you check if this is good?

[am009]

@am009 I removed it can you check if this is good?

One issue is that there may be an inconsistency between the post object and the updated base. the post is updated in every loop. When the base is updated, the loop goes on, and the post object no longer originates from the same try_transform_into call as the base.

I debugged the baby_fuzzer_minimizing. I found that the fuzzer.process_execution always returns None for the corpus_idx. Because the fuzzer's crash or coverage feedback is set to empty (). The tmin stage creates its own feedback (self.create_feedback) to check if the input still triggers the crash or has the same execution path. I got the feeling that We just don't need to call fuzzer.process_execution at all.

[tokatoka]

            let (untransformed, post) = transformed.try_transform_into(state)?;

i'm updating post

[tokatoka]

i don't understand why it run execute() then later calls process_execution() for the second time. it's doing the same thing twice

[am009]

            let (untransformed, post) = transformed.try_transform_into(state)?;

i'm updating post

Let me explains it further. For example, num is two (we run the mutation loop 2 times).

• In the first loop, try_transform_into returns untransformed@1 and post@1,
  • the untransformed passed all checks and untransformed is assigned to base.
• In the next loop, try_transform_into returns untransformed@2 and post@2,
  • this time the untransformed@2 input is not assigned to base.
  • post is updated to post@2.
• loop reached exit condition, and the loop returned the variable post, which is of value post@2
• finally, new testcase is constructed from untransformed@1, but [email protected]_exec(state, new_corpus_id) is called to update metadata for the new testcase. Then the incorrect metadata is added to the testcase.

But still, this is a helpful new approach to prevent the clone.

i don't understand why it run execute() then later calls process_execution() for the second time. it's doing the same thing twice

I think the process_execution came from the original mutational stage.

• Fuzzer run fuzzer.execute_input to execute the input
• Fuzzer run fuzzer.process_execution to use the feedback to check if it is a solution or an interesting input. it returns a corpus id if the input is interesting.

However, the tmin stage does not rely on the feedback of fuzzer, but it has its own feedback (self.create_feedback) to check if the minimized input is still the same.

[tokatoka]

Then the incorrect metadata is added to the testcase.

Ok I see. Then it's fine with your code.

[tokatoka]

I debugged the baby_fuzzer_minimizing. I found that the fuzzer.process_execution always returns None for the corpus_idx. Because the fuzzer's crash or coverage feedback is set to empty (). The tmin stage creates its own feedback (self.create_feedback) to check if the input still triggers the crash or has the same execution path. I got the feeling that We just don't need to call fuzzer.process_execution at all.

I thought this fuzzer.process_execution()'s purpose is to make sure that we don't actually increase the corpus. we process the execution then

                if state.corpus().count() == corpus_count
                    && state.solutions().count() == solution_count

we can compare this later.

Because the fuzzer's crash or coverage feedback is set to empty ()

https://github.com/AFLplusplus/LibAFL/blob/main/fuzzers/baby_fuzzer_minimizing/src/main.rs#L42
It's this one right?

[am009]

In the baby_fuzzer_minimizing example, there are multiple stages.

• First, The code fuzz the harness to find at least crash.
• Then it creates a new fuzzer for minimization, loads initial input from the crash folder, and sets the current corpus to 0, the first crash.
• Finally, it launches a new minimization stage.

I think the feedback you referenced is for the first fuzzing step, not for the minimization stage.

I think it is here, it sets the objective and feedback to (). and the executor's observer is also set to ().

LibAFL/fuzzers/baby_fuzzer_minimizing/src/main.rs

Lines 134 to 137 in f0ee6e0

	let mut fuzzer = StdFuzzer::new(scheduler, (), ());
	// Create the executor for an in-process function with just one observer
	let mut executor = InProcessExecutor::new(&mut harness, (), &mut fuzzer, &mut state, &mut mgr)?;

[tokatoka]

ok i see.
i think it's fine to have that here.
for that specific example, yes process_execution is doing nothing, but maybe for others, it is necessary to evaluate it and then make sure there's no addition to the corpus

[tokatoka]

Looks good thank you 👍