Dyninst: New package

[eschnett]

No description provided.

[eschnett]

While the build works fine, the checks afterwards crash with

[ Info: Checking shared library lib/libcommon.so.12.3.0
  | free(): invalid pointer
  |  
signal (6): Aborted

(libcommon is one of the libraries that's built.) How would I debug this?

[giordano]

It's the dlopening that's failing, maybe the init function of the library does something nasty?

[eschnett]

@giordano I can successfully dlopen each of the shared libraries from within the build script. However, this still fails from within Julia. Do you have an idea why this might be?

I also looked at the libraries' initialization routines but that's a bit hopeless because these are large C++ libraries and I don't know which objects would be created or which functions would be called.

[giordano]

To leave some breadcrumbs, the init function of the shared library is marked with __attribute__((constructor)) and it's called runDYNINSTBaseInit: https://github.com/dyninst/dyninst/blob/ddd2315b562efdca3fe6aa25d6da2a3842c293fa/dyninstAPI_RT/src/RTlinux.c#L644-L661. But this function only calls the function r_debugCheck() and DYNINSTBaseInit: the former looks relatively innocuous (only running an assert: https://github.com/dyninst/dyninst/blob/ddd2315b562efdca3fe6aa25d6da2a3842c293fa/dyninstAPI_RT/src/RTlinux.c#L412), but the latter is more convoluted because it then calls other functions: https://github.com/dyninst/dyninst/blob/ddd2315b562efdca3fe6aa25d6da2a3842c293fa/dyninstAPI_RT/src/RTcommon.c#L163-L174

[giordano]

This is what we do during audit to dlopen the library: https://github.com/JuliaPackaging/BinaryBuilder.jl/blob/54284653349b50eef5ad50b0c579659c38ef99e6/src/Auditor.jl#L150-L159. Not really much (using Libdl; dlopen(/path/to/library))

[eschnett]

I saw the call in Auditor.jl, and I am using the same flags when calling dlopen from C in the build script, which is working (no segfault). I think I don't understand the difference between the build environment and the environment in which the auditor is running in CI. I assume the build script is running in a Docker container that (somehow) makes things work? Maybe a system library is different?

[eschnett]

There are different libstdc++ libraries:

libstdc++.so.6 => /workspace/destdir/lib/libstdc++.so.6 (0x00007fa5b1617000)
libstdc++.so.6 => /usr/lib/csl-glibc-x86_64/libstdc++.so.6 (0x00007f99a555a000)

The first is used by Dyninst, the second by Boost. That doesn't look healthy.

[eschnett]

While the dlopen in the auditor fails, I can:

• generate a package locally
• using Dyninst_jll
• dlopen(Dyninst_jll.libcommon)
  and this does not segfault. I think this is somehow a problem in the auditor.

[eschnett]

@giordano I disabled the auditor. I think (see above) that the generated shared libraries are actually working fine.

[giordano]

I'm not incredibly happy about disabling the auditor, it does more stuff than just dlopening the library, like ensure the runpath is set correctly, the soname is set, etc...

[giordano]

I'm still convinced they do something dodgy in their init function: #7473 (comment). If we have problems during audit here, we can have problems elsewhere, if dlopen is failing we should fix it, rather ignoring it.

[eschnett]

The current options actually only set dont_dlopen, they don't explicitly disable the auditor. I assume this means that the auditor is still running some other tests?

Apart from setting a few global variables Dyninst installs a signal handler in DYNINSTinitializeTrapHandler:

int DYNINSTinitializeTrapHandler(void)
{
   int result;
   struct sigaction new_handler;
   int signo = SIGTRAP;

   // If environment variable DYNINST_SIGNAL_TRAMPOLINE_SIGILL is set,
   // we use SIGILL as the signal for signal trampoline.
   // The mutatee has to be generated with DYNINST_SIGNAL_TRAMPOLINE_SIGILL set
   // so that the mutator will generates illegal instructions as trampolines.
   if (getenv("DYNINST_SIGNAL_TRAMPOLINE_SIGILL")) {
      signo = SIGILL;
   }

   new_handler.sa_sigaction = dyninstTrapHandler;
   //new_handler.sa_restorer = NULL; obsolete
   sigemptyset(&new_handler.sa_mask);
   new_handler.sa_flags = SA_SIGINFO | SA_NODEFER;
   
   result = sigaction(signo, &new_handler, NULL);
   return (result == 0) ? 1 /*Success*/ : 0 /*Fail*/ ;
}

Maybe Julia doesn't like SIGTRAP being caught?