error in 1.131.x crashes go process

[aviramha]

User sees

2025-01-28T19:12:11.654285Z ERROR ThreadId(02) mirrord_layer::error: Error occured in Layer >> ResponseError(NotImplemented)

panic: runtime error: invalid memory address or nil pointer dereference
[signal SIGSEGV: segmentation violation code=0x1 addr=0x0 pc=0xd30786]

go 1.23.2 - running on arm64 macOS inside Docker

[linear]

MBE-690 error in 1.131.x crashes go process

[aviramha]

I'm not sure it solves users' issue yet

[aviramha]

My fix didn't fix 😒

[aviramha]

I believe the issue is that go calls the syscall statfs which uses the statfs64 struct in libc, and we use the statfs struct which is backward compatible, hence when coming from go we should actually call statfs64_detour instead.
this is probably relevant to other syscalls that were added as part of the PR (I think there are 2 structs/functions?)

[t4lz]

Is there a reproduction you guys worked with?

[aviramha]

I couldn't reproduce, but I am pretty sure that if you write a go app that calls statfs when using mirrord it'd get wrong data in the last fields, and that's how we can add some verification to it.

[aviramha]

(On Linux)

[t4lz]

We (implicitly) use the statfs64 struct in the agent when we do the call (we call nix::sys::statfs::statfs and on linux that just calls statfs64), and when we fill the struct in the layer the memory is the same as in the 64 bit struct. However we just don't send over some fields, and instead set them to 0 in the user application's struct.
I'm adding a new protocol message so that we can send over all of the fields.
I don't know if that will prevent the crash, as we don't know for sure the missing fields were the cause for the crash (note the memory of the missing fields was not uninitialized, it was zeroed).

[aviramha]

Still happens after rollback of Go hooks

[aviramha]

It also happens on macOS natively :|