net: failures with signal arrived during cgo execution

[bcmills]

#!watchflakes
post <- pkg == "net" && `signal arrived during cgo execution`

[bcmills]

One of these was reported in #27992 (comment), but doesn't match the prior failure mode for which that issue was created.

I saw another in a TryBot in https://storage.googleapis.com/go-build-log/55480854/freebsd-amd64-12_3_a94bfe2e.log.

[mknyszek]

Yeah, that definitely looks different from #27992, but I'm not sure this is necessarily going to be a C&RT issue? It seems like there's a segfault in C code during a net package test. signal arrived during cgo execution is just indicating that the crash happened in cgo, so the stack is going to be truncated. Tentatively removing the compiler/runtime label, but feel free to add it back if you disagree.

[bcmills]

signal arrived during cgo execution is just indicating that the crash happened in cgo, so the stack is going to be truncated.

Ah, I see. That makes it tricky to track down the actual failure, though — is there some way we can provide the equivalent of a default runtime.SetCgoTraceback for the cgo dependencies in the standard library? (Can we make assumptions about system libraries that allow for something simpler than the full generality of SetCgoTraceback that we need for user C code?)

(CC @ianlancetaylor)

[ianlancetaylor]

If anything, system libraries are harder to get a traceback from, because they are always heavily optimized and because often the debug info is stored somewhere else. (Without the debug info the traceback is close to useless, it's just a list of PC values.)

For debugging purposes a blank import of github.com/ianlancetaylor/cgosymbolizer will often get a C backtrace, but I really can't recommend making that part of the Go standard library. It's 17,000 lines of C code.