On 3/13/18 10:54 PM, David Holmes wrote:
Hi Daniil,

On 14/03/2018 3:26 PM, Daniil Titov wrote:
Please review the changes that fix intermittent timeout failure of serviceability/dcmd/framework/* tests.

The problem here is that these tests invoke jcmd in different ways and one of such ways is when a main class is passed to the jcmd as a VM identifier. The main class for jtreg test is com.sun.javatest.regtest.agent.MainWrapper and in some cases more than one test are running in parallel and there are multiple Java processes with com.sun.javatest.regtest.agent.MainWrapper as a main class . When it happens jcmd iterates over all Java processes that match the condition (the main class equals to com.sun.javatest.regtest.agent.MainWrapper) and executes the command for each of them. That results in the jcmd invokes the given command multiple times and attaches to Java processes not related to the current test.

It's good to finally find the root cause of the problem!

The fix makes serviceability/dcmd/framework/* tests non-concurrent to ensure that they don't interact with other tests.

This seems more of a workaround than a fix - though I don't know whether there is a way to distinguish multiple VMs all running what appears to be the same main class.
The three tests are all intentionally testing the "jcmd <classname>" functionality. A better fix (but not worth it IMHO) would be to spawn a separate test process rather than using the main test process, which is always going to share the main class name with other concurrently running tests.

My concern with the fix is how long it will take to run these tests sequentially, as they are run in tier2 and as part of the CI test job?
They are quick tests and there are only 3 of them. They appear to take less than 5 seconds each.



Bug: https://bugs.openjdk.java.net/browse/JDK-8166642
Webrev: http://cr.openjdk.java.net/~dtitov/8166642/webrev.01

The tests ran successfully with Mach5.

Best regards,

Reply via email to