RFR JDK-7092821 "java.security.Provider.getService() is synchronized and became scalability bottleneck"

[Valerie Peng]

Hi, Max,

In  my first prototype, I got rid of the Entries classes (for both SUN 
and SunRsaSign provider) and adds the Services inside the Provider. 
Later, when I saw VerificationProvider class and its dependency on the 
Entries class, I have to revert back to the original approach of having 
the Entries classes. I thought about putting the Entries classes into 
their provider classes, however, I am not sure if any of our licensee 
may remove SUN, SunRsaSign provider class which is why we have 
VerificationProvider. So, I think having separate Entries classes are 
probably the safest choice. As for creating a HashSet to store the 
services and then putting them into provider object in individual 
provider class is to work with the current API of Provider.putService()...

If you have better suggestion, I am all ears...

Thanks!
Valerie


On 12/3/2018 7:57 PM, Weijun Wang wrote:
> In fact, if the only reason why those Entries classes exist is because of VerificationProvider, can we move the content back into their Provider class as a public static method which is still callable by VerificationProvider?
>
> —Max
>
>> 在 2018年12月4日，11:43，Weijun Wang <weijun.wang at oracle.com> 写道：
>>
>> Hi Valerie
>>
>> I'm looking at the put->putService changes. Before this, a single put() in SunRsaSignEntries call adds a service but now you need to add() into a LinkedHashSet first in SunRsaSignEntries and then putService() each back in SunRsaSign.
>>
>> I don't have a simple way to do this since Provider::putService is protected and SunRsaSignEntries is also used in VerificationProvider, just wonder if you have thought about this.
>>
>> Thanks
>> Max
>>
>>> On Nov 22, 2018, at 2:05 AM, Valerie Peng <valerie.peng at oracle.com> wrote:
>>>
>>> Hi,
>>>
>>> Can someone help reviewing this fix?
>>>
>>> Besides changing the Provider class to use ConcurrentHashMap in order to reduce the lock contention on Provider.getService() calls, I also changed the security providers in java.base module to register through putService(...) calls. Performance is manually verified and mach5 run is clean.
>>>
>>> Bug: https://bugs.openjdk.java.net/browse/JDK-7092821
>>> Webrev: http://cr.openjdk.java.net/~valeriep/7092821/webrev.00/
>>>
>>> Thanks,
>>> Valerie