Code Review Request JDK-8170329 New SSLSocket testing template

[Xue-Lei Fan]

On 12/2/2016 4:34 PM, Artem Smotrakov wrote:
> Hi Xuelei,
>
> Please see inline.
>
> On 12/02/2016 03:53 PM, Xue-Lei Fan wrote:
>>> Let's whether Sean or Weijun can have free cycle for the review of
>>> this part.
> Yeah, that would be great.
>>
>>> - Why did you remove Peer and Application interfaces? I think those
>>> interfaces make SSLSocketTemplate more flexible since it allows override
>>> doServerSide/doClientSide logic if necessary - it doesn't seem to be
>>> worse. If there is no strong reason to remove those interfaces, I would
>>> prefer to keep them with default static/stateless
>>> doServerSide/doClientSide versions.
>> I agree with you that The Peer and Application interfaces are more
>> flexible.  I have no strong reason to remove them.  For this update, I
>> focus more on simple and minimal sub-classes.  The Peer and
>> Application interfaces need some complicated coding (condition,
>> threading, etc) in sub-classes, so the design does not fall into the
>> scope.  I may prefer to remove them at this update and see what
>> happens if we moving forward with more test update. We can add them
>> back or create a more flexible one if we need the flexibility in the
>> future.
> We have a lot of JSSE tests which may need to do something else in
> doServerSide() and doClientSide() methods (that's what I learned from my
> experience working with JSSE tests). Peer and Application interfaces
> basically allow to override doServerSide() and doClientSide() methods.
> Your changes make doServerSide() and doClientSide() methods private, so
> that test can't add something inside it.
>
> If you think those interfaces are helpful, I think it's better to keep
> Peer and Application interfaces instead of removing/adding them back and
> forth.
>
> Another option may be providing methods which are called in
> doServerSide() and doClientSide(). Those methods may be overridden by a
> test if it needs to do extra stuff inside doServerSide() and
> doClientSide().
>
> I am fine with both ways, so it's up to you. But I would like to make it
> clear what way we want to follow.
>
> What do you think?
You points above are good.  When I made the update, I struggled a lot 
about the balance of flexible and simplicity.  I was inclined to 
simplicity at last as we can update the template if any new feature is 
required in the future.  I'm not sure how much this template can be 
used, exposing as less methods as possible will give us more flexible to 
make further update.  So I made most of the methods private, and removed 
a lot of methods.  Let's update the access scope if we really need them 
in the future.  A small public footprint is easier to start.

>>
>>> It also might make sense to make
>>> createSSLContext() a part of public API which I think would make the
>>> template more flexible.
>> We have had the protected createClient/ServerSSLContext() methods.
> Making createSSLContext() accessible doesn't seem to be worse to me. It
> may be final if you don't want anyone to override it for some reason.
>>
>>> - Exceptions are printed out in startServer/startClient methods, it
>>> doesn't look necessary to use suppressed exceptions and initCause()
>>> method. What was wrong with the code in runTest() method? The code in
>>> runTest() method looks shorter and cleaner to me.
>>>
>> The main issue of runTest() is that it does not throw the original
>> exception.  Exception tacks have more information for debugging. I
>> want to keep the good side of Brad's previous hardness on this point.
> It is not clear what's the original exception should be because you have
> client and server. If you print all exceptions in
> startServer/startClient (right after they occurred), then you don't hide
> anything helpful for debugging.
>
The exception dump may have more information than the single exception 
track, for example the nested cause stacks.  The original exception 
handling was very simple, and a lot improvements were made so that the 
log is easier for debugging.

> IMO, suppressed exceptions may confuse here (that's just from my
> experience looking at JSSE test logs). The logic in lines 739-763
> doesn't make it cleaner what exception was thrown on what side. It does
> "local.initCause(remote)", but actually "remote" is not a cause of
> "local". Finally, it does "exception.addSuppressed(startException)"
> where "startException" can be thrown either on server or client side
> which actually depends on test logic (an engineer needs to keep it in
> mind).
>
> Would it be better to have in logs something simple like the following?
>
> ....
> This is an unexpected exception on client side:
> <full stack trace>
> This is an unexpected exception on server side:
> <full stack trace>
> Test failed.
> ...
>
> Doesn't it look simpler?
The purpose of suppressed exception is fully use of the exception for 
debugging.  The above suggestion still lose some exception debugging 
information.  I dumped the exception message (line 808/818), hopefully 
it can help a little bit if confusing.

It's good suggestion that we'd better to tell which side (the client or 
server side) throws the exception.  I will think more about it in the 
next webrev update.

>>
>>> - lines 114-123, this code is used quite often by tests, why don't we
>>> keep it in SSLSocketTemplate?
>>>
>> Good point!  I don't like it in SSLSocketTemplate as it is used for
>> HTTP connections only, but it may be worthy to have a sub-template for
>> HTTPS client testing.  What do you think if we address this
>> enhancement in a new bug?
> I am fine with it. My point is that SSLSocketTemplate may contain some
> useful static methods like this (to avoid duplicate code).
>
I see.  I will consider it more in the next webrev.

Thanks,
Xuelei