Call Frame Information (CFI) and frame pointer considerations

[briansmith]

Could you please expand the documentation regarding ABI compatibility with respect to call frame information (CFI) and frame pointers? In particular, are there no CFI directives in the assembly because they wouldn't be an improvement, or because it hasn't been a priority/desire to add them? This is something that comes up when comparing your assembly language code to other implementations like OpenSSL's and BoringSSL's.

See https://www.imperialviolet.org/2017/01/18/cfi.html for more clarification about what I'm referring to.

Thanks in advance.

[jargh]

This is mostly just something we haven't prioritized so far, though I remember someone mentioning them as desirable before (perhaps it was Adam Langley). We pushed it to the background as something that "only matters for debugging", but I admit it's not ideal.

There is currently no attempt to maintain actual frame pointers in the code, and it would be disruptive to do so. Quite a lot of the x86 functions (pretty much all the big ones I think) modify RBP internally; a few of the ARM ones also modify X29, though in both cases the value is saved and restored as per ABI.

However, my understanding is that we could indeed provide suitable CFI directives, since no text/data segment changes (and hence no proof changes) would be needed. Given that it would be much better for debugging or related activities, this indeed seems like something we should do. We'd better take the time to make sure we know what we're doing though :-)

[briansmith]

It isn't really about debugging. Many environments, maybe even AWS internally, require the call frame information so they can do continuous live profiling of production systems, or at the very least on-demand profiling. The CFI directives help avoid the disruption caused by using ebp for a purpose other than the stack frame pointer.

[jargh]

You're right of course, so this seems to be something we should do. If we make a first attempt, would you be willing to glance over it and see that it looks reasonable?

Also, do you have any thoughts on practical validation? I guess it could be sanity-checked by running in some simulated environment like valgrind or gdb and deliberately putting breakpoints in, somehow or other. Do you know of any established methodology used by BoringSSL or anyone else?

[briansmith]

One way to validate it would be to write C functions with the same signature (parameters and return value) as every function you define in assembly, with a dummy body like this:

{
dummy();
}

__attribute(noinline)__
void dummy() { }

For each function, change the call to dummy with some minimal approximation of whatever you do, alloca, etc. Then you can diff the assembly generated by the C compiler against what you are doing.

Of course, this is cringeworthy compared to your rigorous validation of correctness for your code. But, it may be a good first step towards eventually incorporating the validation of CFI into your proofs.

[briansmith]

Check out BoringSSL's CFI testing framework: https://issues.chromium.org/issues/42290050

[jargh]

There is a first attempt at adding CFI here: #214. If you have time I'd appreciate a quick glance to see if it looks reasonable. So far there has been so serious validation, but the text section of the code is unchanged and hence all proofs still work.

[briansmith]

Just reading https://github.com/torvalds/linux/blob/master/tools/objtool/Documentation/objtool.txt I see three things that are potentially useful here:

1. The objtool tool itself may be useful for verifying that the CFI annotations are correct. My understanding is that it is designed to support non-Kernel use cases, though I see it mostly focuses on the Linux Kernel's special "orc" unwinding.
2. The document describes the rules for CFI annotations.
3. The document mentions some macros that the Linux kernel assembly code uses, which seem like they would be useful here, though I guess maybe license issues may get in the way.

[briansmith]

https://maskray.me/blog/2020-11-08-stack-unwinding is another useful resource.