multiple-debuggers test: testprog is left running when test times out

[llvmbot]

Bugzilla Link	25063
Resolution	FIXED
Resolved on	Oct 06, 2015 10:44
Version	unspecified
OS	All
Reporter	LLVM Bugzilla Contributor
CC	@labath

Extended Description

When the MultipleDebuggers test times out, I'm seeing test/api/multiple-debuggers/testprog programs sticking around.

Figure out why that is. This might be a corner case of the test system sending a soft terminate to the parent, which might kill the parent, but doesn't guarantee it kills the children of the parent process.

We might need/want to send a process group hard terminate after we wrap up to ensure all children are terminated even if we think the main process is done.

[labath]

lldb starts inferior processes in a separate process group (at least on linux, but I'm pretty sure its OSX as well) precisely to avoid the thing we do on test shutdown: having signals delivered to the debugger make their way to the inferior.

Unfortunately this means cleaning up these processes is nearly impossible. (This wasn't the case for TestAttachDenied, because here the process was started from the test suite.) I have had some ideas on how to handle this, but I did not deem it worth the trouble. E.g. on linux we could use PR_SET_CHILD_SUBREAPER (prclt(2)), but I don't know if there are similar functionalities on other OSs.

[llvmbot]

On the OS X side, the issue is that we start the processes suspended, which puts them in a state that the kernel will absolutely refuse to service signals for --- it is really not running yet.

I thought about adding a clean-up step in the test runner that scans for process group members that are not yet dead at test exit, and if a hard kill doesn't take care of them, then use the in-test-runner-process lldb module to attach and then kill, since this is all that is really needed. But rather than add that complication, I went with taking care of the core issue, which is doing too much all at once in a parallel test runner environment. I could have tried to calculate the number based on # cores, but that's not really good enough. So for now I cut it from 50 to 20 after talking to Jason Molenda about it. This is plenty to stress what was desired to be stressed at the time the test was written.

Change to # concurrent debuggers made here:

svn commit
Sending test/api/multiple-debuggers/multi-process-driver.cpp
Transmitting file data .
Committed revision 249421.