JAXP default implementation and JDK-8152063

[huizhe wang]

On 3/29/2016 12:21 AM, Alan Bateman wrote:
>
> On 28/03/2016 23:46, huizhe wang wrote:
>> Thanks David.  So I understand the dynamic nature of the server 
>> configuration. There maybe two options to solve it:
>>
>> 1) Add a system property to point to a Layer in order to find an 
>> alternative-system-default. This will add a new step to the JAXP 
>> process after the current ServiceLoader process.  I saw that you had 
>> concern over the performance of searching a provider among all 
>> modules in a Layer.
> I'm not sure that it would be feasible to statically configure the VM 
> to use a module in a layer of modules that will get created sometime 
> later in the lifetime of the VM. Also this amounts to changing the 
> default implementation on the fly and something we should avoid 
> without deep consideration.

This option is similar to that previously suggested using Layer, except 
it doesn't require a new method that takes a Layer. I thought a layer 
would have already been created after the container had loaded its 
modules. So, would the container pre-load its modules including the one 
that might have been configured with a JAXP provider?
>
> On the performance issue, this is a temporary issue with a new API 
> where we had to remove indexing in order to make progress. It 
> shouldn't impact anything and will sort it out once we get further on 
> a number of topics.
>
>>
>>
>>
>> 2) Add a new type FinderDelegate for processes such as the "proxy" in 
>> your case to implement. If the FinderDelegate process fails to locate 
>> a provider, it would signal the jaxp process (by returning null) to 
>> fall back to the JDK-default implementation. In other words, when the 
>> system property points to a FinderDelegate, the 4-step JAXP process 
>> is reduced to two: delegate the process to the FinderDelegate, and 
>> fall back to the system default implementation.
> The devil is in the detail of course. You haven't said if the 
> FinderDelegate implementation has to be visible via the system class 
> loader.

No, not the JDK system class loader, but probably the container's system 
or bootstrap loader. It belongs to the container, same as the "proxy" 
David mentioned in the original email. JAXP would try to load it the 
same way as it would with an implementation class.

>
> I think the main thing is to tread carefully and it would be very easy 
> to introduce a troublesome mis-feature here.

Sure, we'd have some review circles to go through.

-Joe

>
> -Alan.