code review request: 7162687: enhance KDC server availability detection

[Weijun Wang]

On 05/23/2012 09:25 AM, Valerie (Yu-Ching) Peng wrote:
>
> The source changes look fine. The re-trying receive() part seems not
> useful, but if you want to be conservative, that's fine with me too.
>
> What is the following change for?
>
> @compile -XDignore.symbol.file MaxRetries.java

With this line, the compile won't show so many warnings on sun.* class 
reference. It's not really necessary, but I want to jtr to be cleaner.

>
> BTW, there is typo on line 43, "tp" should be "to"

Oh.

>
>> PortUnreachableException is not available on Mac, new test won't work
>> there.
> So, what is the behavior on Mac then? Is it a different type of
> SocketException or else?

On Mac, connect() does not call the native connect so PUE is not thrown. 
It simply waits until a SocketTimeoutException is thrown. The reason is 
7122794.

Thanks
Max

>
> Thanks,
> Valerie
>
> On 05/18/12 03:08, Weijun Wang wrote:
>> Hi Valerie
>>
>> Webrev updated:
>>
>> http://cr.openjdk.java.net/~weijun/7162687/webrev.01/
>>
>> Summary:
>>
>> KdcComm.java and NetClient.java
>>
>> 1. Make NetClient AutoCloseable, use try-with-resources.
>>
>> 2. Add a connect() after UDP socket is created, this ensures an
>> immediate PortUnreachableException if the server is not on. When this
>> exception is seen, we won't try the same KDC again.
>>
>> 3. Now we have
>>
>> try {
>> dgSocket.receive(dgPacketIn);
>> }
>> catch (SocketException e) {
>> + if (e instanceof PortUnreachableException) {
>> + throw e;
>> + }
>> dgSocket.receive(dgPacketIn);
>> }
>>
>> Here I have no idea what benefit the original code gets by retrying
>> receive(). This code has been there at the very beginning. I simply
>> add a new exception check.
>>
>> test/sun/security/krb5/auto/Unreachable.java
>> test/sun/security/krb5/auto/unreachable.krb5.conf
>>
>> New test. With PortUnreachableException, the test should finish in a
>> very short time without waiting for 3 timeouts. Hopefully no process
>> is waiting on those 3 ports I hard code.
>>
>> test/ProblemList.txt
>>
>> PortUnreachableException is not available on Mac, new test won't work
>> there.
>>
>> test/sun/security/krb5/auto/BadKdc.java
>> test/sun/security/krb5/auto/MaxRetries.java
>>
>> Old tests need to be updated. Start a dummy DatagramSocket to trigger
>> the TimeoutException.
>>
>> test/sun/security/krb5/auto/TcpTimeout.java
>>
>> The test checks for time elapse but quite a lot of our test machines
>> are virtual machines and Windows guests always get timing wrong.
>> Remove that check.
>>
>> Thanks
>> Max
>>
>>
>> On 04/19/2012 11:21 AM, Weijun Wang wrote:
>>> Hi Valerie
>>>
>>> Please take a review at
>>>
>>> http://cr.openjdk.java.net/~weijun/7162687/webrev.00/
>>>
>>> Basically I add a connect() before send/receive and if a KDC UDP server
>>> is not on, an ICMP Port Unreachable message will be received
>>> immediately, and there will be no more retry on the same KDC.
>>>
>>> In the tests, an idle DatagramSocket must be created to avoid the ICMP
>>> message being sent. I also enhance the BadKdc tests to check the new
>>> behavior: When no idle DatagramSocket is created, there are no more
>>> retries on the same KDC, but you can still see krb5.kdc.bad.policy
>>> working.
>>>
>>> Thanks
>>> Max
>>>
>>>
>>> -------- Original Message --------
>>> *Change Request ID*: 7162687
>>> *Synopsis*: enhance KDC server availability detection
>>>
>>> Product: java
>>> Category: java
>>> Subcategory: classes_security
>>> Type: RFE
>>>
>>> === *Description*
>>> ============================================================
>>> Currently if there are multiple KDCs defined for a realm, we send UDP
>>> request to each one and wait for reply, when no reply is received, we
>>> retry. After several unsuccessful retries, we go on to the next KDC. The
>>> timeout for the wait is normally 30 seconds. If the KDCs defined are not
>>> all turned on, the time spent to finally get a response might be very
>>> long. This is especially true for customers having a main KDC and
>>> several slave KDCs and the main KDC is in maintenance.
>>>
>>> In fact, there is a better way to detect if the KDC server is on by
>>> connect() to it first. In this way, an IMCP PortUnreachableException
>>> will be thrown in a later send/receive method immediately. Also, when
>>> this exception is thrown, we can be sure that the KDC server is off and
>>> there is no need to retry this one.
>>>
>>> Please note that if a KDC is on but for various reasons it cannot reply
>>> in a timely manner, we will still wait for the timeout and do a retry.
>>>
>>> *** (#1 of 1): 2012-04-19 02:56:04 GMT+00:00 weijun.wang at oracle.com
>>> *** Last Edit: 2012-04-19 02:57:44 GMT+00:00 weijun.wang at oracle.com
>>>
>