test: in coverage runs, log stacktraces directly to stderr

[jmarantz]

Description:
Coverage tests run with -l trace --log-path /dev/null, in order to ensure that all of the code-paths from the maximum level of tracing are covered in tests, but we don't wind up filling up CI with useless detailed artifacts.

The downside of this is that if there's a crash, the backtrace is lost, as the backtracing mechanism uses logging. This PR resolves this by redirecting crash logs to stderr for coverage tests only.

Forcing a segv in a test, I see this from coverage with this PR:

[ RUN      ] IpVersions/MainCommonTest.ConstructDestructHotRestartDisabledNoInit/IPv4
#0 Envoy::SignalAction::sigHandler() [0xfc1a564]
#1 __restore_rt [0x7f3da788f390]
#2 __llvm_coverage_mapping [0x53f0501]
#3 __llvm_coverage_mapping [0x53f02f9]
#4 std::__1::operator<< <>() [0x542c838]
#5 Envoy::MainCommonBase::run() [0x7b84ef0]
#6 Envoy::MainCommon::run() [0x7b63693]
#7 Envoy::MainCommonTest_ConstructDestructHotRestartDisabledNoInit_Test::TestBody() [0x7b5a9cf]
#8 testing::internal::HandleSehExceptionsInMethodIfSupported<>() [0x10fabfc6]
#9 testing::internal::HandleExceptionsInMethodIfSupported<>() [0x10f98501]
#10 testing::Test::Run() [0x10f82dd2]
#11 testing::TestInfo::Run() [0x10f838d8]
#12 testing::TestSuite::Run() [0x10f84049]
#13 testing::internal::UnitTestImpl::RunAllTests() [0x10f91056]
#14 testing::internal::HandleSehExceptionsInMethodIfSupported<>() [0x10fb0e66]
#15 testing::internal::HandleExceptionsInMethodIfSupported<>() [0x10f9b321]
#16 testing::UnitTest::Run() [0x10f90aa0]
#17 RUN_ALL_TESTS() [0xf6bc105]
#18 Envoy::TestRunner::RunTests() [0xf6bb8db]
#19 main [0xf6b9f98]
#20 __libc_start_main [0x7f3da74d4830]
external/bazel_tools/tools/test/collect_coverage.sh: line 128: 17456 Segmentation fault      (core dumped) "$@"

Note that this was tested manually; we do not have a test for this test behavior.

Risk Level: medium -- mostly because this messes with the crash handler
Testing: //test/..., plus a manual coverage run with a segv injected.
Docs Changes: n/a -- this does not add an option for envoy servers; just for tests.
Release Notes: n/a

[mattklein123]

Thanks for working on this. This is going to be super useful in figuring out our coverage issues. A few small comments.

/wait

[mattklein123]

Should we just use TCLAP here? It seems like we are reimplementing a bunch org arg parsing for tests. Also, is there any reason to make this a real server option with documentation? Probably not but just throwing it out there.

[jmarantz]

We could; actually I have that in a patch saved away if we want to go in that direction.

OTOH we could use the TCLAP library to do this parsing rather than do this manually; I just copied the loop from above. I don't know if we can replace the one above because it uses the --name=value syntax rather than --name value, so I was trying to use a parallel approach but I'm not wedded to it.

[jmarantz]

actually there's a third option I considered, using #ifdef ENVOY_CONFIG_COVERAGE and not making it a command-line option at all. WDYT?

[mattklein123]

+1 for simple. :)

[mattklein123]

(I think you would have to ifdef out that test which checks the log but that is probably fine)

[mattklein123]

nit: prepend with BackwardsTrace or something?

[jmarantz]

will do; or I could just delete it.

[mattklein123]

Sure either way

[mattklein123]

Awesome, thank you!!!

[jmarantz]

Sorry can you re-approve? Formatting issue needed to be fixed.