Inconsistent and mostly undocumented default exit codes

[haasn]

Turns out my CI was completely ignoring msan failures for the longest time because an exit code of 77 (msan's default exitcode) is interpreted as "skip this test" in the GNU testing protocol (used by meson etc.).

I found it essentially entirely impossible to figure out this fact from the documentation alone, I was only able to find SanitizerCommonFlags which claims that exitcode defaults to 1, as well as various other bits of documentation that don't even mention the exit code.

But looking at the source code of rt-lib, the actual behavior of the exit code seems a bit sporadic:

• msan defaults to 77 (msan/msan.cpp)
• asan defaults to 1 (asan/asan_flags.cpp)
• ubsan is hard-coded to Die() (which implies exit(1)), and ignores cf.exitcode entirely!
• tsan defaults to 66 (tsan/rtl/tsan_flags.cpp)
• lsan defaults to 23 (lsan/lsan.cpp)
• hwasan defaults to 99 (hwasan/hwasan.cpp)
• scudo defaults to 1 (scudo/scudo_flags.cpp)
• ...

Quite frankly this needs to be documented more clearly at least somewhere, so that users aren't left grepping the source code to figure out what tools exit with what status codes and how to change them.

As a side note, exiting with 77 definitely should be changed for the reason mentioned, although rather than setting it to something else absurd like 55, I think the behavior and default should be consistent across all of these tools.

[yugr]

+1 for having tools abort instead of exit.

[haasn]

Oh, yes, abort() is also a good idea, specifically because it implies the generation of core dumps which can assist in debugging the issues involved.

[rnk]

A couple things:

• With MSan, the shadow memory typically makes core dumps so large as to be unusable. With effort, the shadow could be unmapped to make the dump smaller, but this is not simple.
• Many users hook abort or have SIGABRT signal handlers. This can make it hard to reliably exit with a clean error. Directly calling the exit syscall is IMO a better default.

But, maybe there is a better exit code than 77, I have no info on why 77 was chosen.

[eugenis]

Yes, we are not being very consistent here. 77 was picked at random (and it also looks nice). I'm not sure about changing the defaults at this point, there could be users who depend on these values.

…

On Wed, Apr 15, 2020 at 8:30 AM Reid Kleckner ***@***.***> wrote: A couple things: - With MSan, the shadow memory typically makes core dumps so large as to be unusable. With effort, the shadow could be unmapped to make the dump smaller, but this is not simple. - Many users hook abort or have SIGABRT signal handlers. This can make it hard to reliably exit with a clean error. Directly calling the exit syscall is IMO a better default. But, maybe there is a better exit code than 77, I have no info on why 77 was chosen. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1225 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AADG4SV6JUAOVUUNWSXX2XDRMXHKXANCNFSM4MIIFILA> .

[Flowdalic]

Related change request https://reviews.llvm.org/D92490