Implementing Data Flow Integrity

Dear all,

I want to implement a pass that provides some kind of data flow
integrity similar to Write Integrity Testing
(https://www.doc.ic.ac.uk/~cristic/papers/wit-sp-ieee-08.pdf).

This approach statically determines for each memory write the
(conservative, overapproximated) points-to set of locations that can be
written by the instruction. Further, it instruments the memory write
instruction to prevent a write to a location not in the points-to set.

How can I get the points-to set, including locations from
stack/heap/static variables?
How do I approach this problem in general?
I am new to LLVM.

Thank you!

Regards,
– Fredi

Dear all,

I want to implement a pass that provides some kind of data flow
integrity similar to Write Integrity Testing
(https://www.doc.ic.ac.uk/~cristic/papers/wit-sp-ieee-08.pdf).

This approach statically determines for each memory write the
(conservative, overapproximated) points-to set of locations that can be
written by the instruction. Further, it instruments the memory write
instruction to prevent a write to a location not in the points-to set.

Correct. I would also point out that their use of Anderson's analysis is (most likely) unnecessary. Because they unify points-to sets before instrumenting, they are modifying the end-result of the inclusion-based analysis to be what unification-based points-to analysis would have computed. It is not clear to me that anything can be gained by using inclusion-based analysis over unification-based analysis.

How can I get the points-to set, including locations from
stack/heap/static variables?
How do I approach this problem in general?

To the best of my knowledge, the existing LLVM alias analysis passes only provide a mod/ref and aliasing query interface. I don't believe they provide a shape graph or points-to sets that can be easily used. You might want to check CFL-AA to see what it provides, but I have personally never used it.

You could use DSA located in the poolalloc project which provides a shape graph. The original SAFECode essentially did what WIT does (except that it also protected memory reads and used a very different run-time check mechanism, plus it could optimize away provably type-safe checks). SAFECode used DSA's shape graphs to segregate the heap, find points-to sets, and learn memory object type information. However, in its current shape, you'd need to run DSA prior to most LLVM optimizations to get good field sensitivity. Otherwise, DSA will lose field sensitivity and provide poor precision in its results.

As I need something similar for my research work, my research group will be working on either improving or replacing DSA. However, it'll be awhile, so if you need something now, either CFL-AA or DSA will be your best bet.

I am new to LLVM.

Welcome to the club.

Regards,

John Criswell

Hi John Criswell,

Thank you for your helpful answer.

I think CFL Alias Analysis is not the right way to go because it seems
to avoid building whole points-to sets to be more efficient (correct me
if i am wrong). So, it would trade compile-time performance for
runtime-security, which is bad.

DSA from poolalloc seems promising, I need to check it.

Do you think that an implementation of Andersen's pointer analysis
(like https://github.com/grievejia/andersen) would work?
It seems to be not field-sensitive. Do you have any clue how hard it is
to make an existing analysis field-sensitive?

Regards
—Fredi

Hi John Criswell,

Thank you for your helpful answer.

I think CFL Alias Analysis is not the right way to go because it seems
to avoid building whole points-to sets to be more efficient (correct me
if i am wrong).

It does not build points-to sets, but does answer all the aliasing queries
you can normally ask of LLVM.
It will answer all of those queries as well as any other anderson's impl
will.

Nothing in LLVM asks or answers "what is the set of things this pointer
points-to".

So, it would trade compile-time performance for

runtime-security, which is bad.

This is not correct, in the sense that CFL-Anders will give precisely the
same answers any other andersen's impl will give about aliasing.

DSA from poolalloc seems promising, I need to check it.

Do you think that an implementation of Andersen's pointer analysis
(like https://github.com/grievejia/andersen) would work?

That one ignores the effect of certain instructions.

It seems to be not field-sensitive. Do you have any clue how hard it is
to make an existing analysis field-sensitive?

For LLVM, it requires changes to both the type of constraints one handles,
constraint building, constraint optimization, and constraint solving.

It is non-trivial unless you are very familiar with andersens and it's
various implementation techniques.