Great, the new busy port algorithm looks better.
The port assigned will always be one that was available and is now busy
to prevent the registry creation.
As expected, there is a small window between the (try-with-finally)
close of the server socket channel
and the 2nd createReg. But that can't be avoided. If the port is
re-used in that gap
the test will fail. (And the exception handler at 77 will see an in-use
67: The exception being caught is already one thrown by TestLibrary.bomb
so it would be
cleaner to just re-throw e; or better yet, just don't bother to catch
That exception should cause the test to fail.
It may be personal coding style but the createReg method with a boolean
flag to suppress
the exception is just more confusing that putting the code in-line in
the two places it is used.
On 4/8/2018 4:10 AM, Hamlin Li wrote:
I have changed to not use RegistryVM at all, please review the patch:
On 04/04/2018 10:15 PM, Roger Riggs wrote:
I think this change to use a separate process for the 2nd registry
changes the test so that it does not
address the original test case. The original problem was the
incorrect retention of an object in
the object table when the create of a registry in the same process
Finally being able to create the registry in the same process assured
that the object was not
retained in the object table.
Go back to creating the 2nd registry in the test process.
RegistryVM.java: 2, the copyright update should be "2017, 2018," range.
(I'm really not a fan of all the RegistryVM methods with the same
name and minimal and implicit
differences in their functions. When reading the test, you have to
go and read the RegistryVM method
to see what it does. I would have preferred that the full
createRegistryVM (out, err, options, port) method
was used directly in the tests. In the case of a method used once,
it is an inconvenience method, not a convenience).
The new terminate() method is quite similar to the existing cleanup()
method which does not wait.
It would be a good cleanup to figure out if another method is really
The override is going to change the behavior of the existing uses of
It should be checked that it does not break any existing uses.
178: 187: The method comments are not consistent, one says forcibly
and the other gracefully
but both call requestExit and both call destroy().
On 4/3/2018 11:35 PM, Hamlin Li wrote:
Hi Joe, Roger,
Thank you for reviewing, I have refactored the test more and fix as
please review http://cr.openjdk.java.net/~mli/8188897/webrev.01/
On 04/04/2018 10:42 AM, Joseph D. Darcy wrote:
Some general comments on the coding for tests like this:
* It is preferable to avoid sleep in tests to avoid increasing the
minimum amount of time a test takes to run. This helps limit the
overall time it takes the test suite to run.
* If timeouts are used, it is recommend to factor the maximum time
waited with the jtreg timeout scaling factor; I don't recall its
exact name. In other words, many of our tests are run on heavily
loaded systems and the tests take longer than run than on typical
laptops and workstations so jtreg is invoked with an timeout
scaling factor. Individual tests should be sensitive to this
scaling factor for any internal timeout that need to be used.
On 4/3/2018 7:48 AM, Roger Riggs wrote:
Instead of a simple time delay, it would be useful to wait for the
RegistryVM to terminate.
In killRegistry: 149, adding subreg.waitFor() should be sufficient.
58: If using a 'for' loop it would be easier to understand if it
included the usual start, increment and termination.
Instead of burying it in the exception handler.
59, 102, 104: the introduction of the kill boolean makes the test
harder to understand and seems to be unnecessary.
the killRegistry() method already will only kill the subprocess
if it still is alive.
On 4/2/2018 6:33 AM, Hamlin Li wrote:
would you please review the following patch?