RTSan unit-tests can crash due to use of hard-coded file descirptors

[jmorse]

Hi -- we (Sony) occasionally see crashes coming out of the RTSan unit tests. The root cause boils down to some of the file-descriptors being passed into syscalls actually being legitimate data under some circumstances. In combination with llvm-lit using GoogleTest sharding, different unit tests are interleaved and we see emergent effects, such as this sequence:

• CloseDiesWhenRealtime [0] closes file descriptor zero -- which is a legitimate file description, it's stdin on posix hosts,
• OpeningASocketDiesWhenRealtime [1] creates a socket, which comes back as file descriptor zero,
• SendToASocketDiesWhenRealtime [2] writes nothingness to file descriptor zero, and the process crashes with SIGPIPE because fd zero is a socket and not connected to anything. It would otherwise receive a EBADF return code. (Edit: actually ENOTSOCK)

These crashes have come and go; currently they're gone because GoogleTest sharding doesn't string together the sequence above, but it might do as more tests are added to the file. I feel to improve test robustness, it would be much safer if all syscalls used a constant illegal file descriptor instead of hardcoding zero. It's also worth noting that the kNotASocketFd constant is zero, a legitimate file descriptor, where it should probably be -1.

[0]

llvm-project/compiler-rt/lib/rtsan/tests/rtsan_test_interceptors_posix.cpp

Line 453 in 7cc4472

auto Func = []() { close(0); };

[1] https://github.com/llvm/llvm-project/blob/7cc4472037b43971bd3ee373fe75b5043f5abca9/compiler-rt/lib/rtsan/tests/rtsan_test_interceptors_posix.cpp#L1345C29-L1345C59
[2] https://github.com/llvm/llvm-project/blob/7cc4472037b43971bd3ee373fe75b5043f5abca9/compiler-rt/lib/rtsan/tests/rtsan_test_interceptors_posix.cpp#L1351C29-L1351C58