bottleneck by java.lang.Class.getAnnotations()

[Peter Levart]

On 11/02/2012 12:04 PM, Peter Levart wrote:
> On 11/02/2012 11:03 AM, Alexander Knöller wrote:
>> Hello there.
>>
>> (Reposting this request for improvement as suggested in mailing list 
>> jdk6-dev)
>>
>> java.lang.Class.getAnnotations() always enters a synchronized-block 
>> (initAnnotationsIfNecessary() ), slowing down multi core machines 
>> that heavily make use of Annotations.
>> (in our Case we use LoadTimeWeaving in the spring-framework 3.1.2)
>> We are using sun-jdk 6u27 on CentOS which has the same 
>> performance-bottleneck. I could not easily find sources for sunjdk 
>> (no open source anymore).
>> So I do not know if Versions 7 or 8 contain fixes for this.
>> openjdk7 and 8 show no fix so far, although it looks like it might be 
>> possible to build a kind of double-checked locking using local 
>> variables?
>>
>>
>> I am not very familiar with concurrency while using SoftReferences, 
>> but I guess using a local Variable for concurrency-avoidance for the 
>> annotations-field or the target of the SoftReference, then doing a 
>> nonsynchronized check for the redefinition and a potential null-value 
>> on the local copy of the annotations-variable should suffice to 
>> decide that one could leave out the synced call and just returns the 
>> annotations-value (referenced by the local variable)?
>> Also you would need to use a local variable while calculating the 
>> annotations prior to assigning the result to the annotations-field to 
>> avoid concurrency-effects on the double-checked locking.
>>
>> Since Code using getAnnotations() always could get hit by an 
>> annotations-result not fitting any more to the class because of a 
>> concurrent thread redefining the class we would not need to take care 
>> of this (and the current code could cause unexpected behaviour 
>> already inside "getAnnotations()" after exiting the lock on 
>> initAnnotationsIfNecessary()).
>>
>> Regards
>> Alex
>>
>> Anfang der weitergeleiteten E-Mail:
>>
>>> Von: Alexander Knöller <alexander.knoeller at gmail.com>
>>> Betreff: Re: bottleneck by java.lang.Class.getAnnotations()
>>> Datum: 2. November 2012 08:43:06 MEZ
>>> An: Joe Darcy <joe.darcy at oracle.com>
>>> Kopie: Alan Bateman <Alan.Bateman at oracle.com>, 
>>> jdk6-dev at openjdk.java.net
>>>
>>> I know. But there is a usual solution to those kind of problems: 
>>> double checked locking.
>>> This would avoid synchronization for all the cases where no 
>>> redefinitions take place.
>>> (I also put in a bug-report for the sunjdk where I elaborated this a 
>>> bit more.)
>>> I am not so familiar with concurrency while using SoftReferences, 
>>> but I guess using a local Variable for concurrency-avoidance for the 
>>> annotations-field or the target of the SoftReference, then doing a 
>>> nonsynchronized check for the redefinition and a potentially 
>>> null-value on the local copy of the annotations-variable should 
>>> suffice to decide that one could leave out the synced call and just 
>>> returns the annotations-value (in the local variable)?
>>>
>>> Also you would need to use a local variable while calculating the 
>>> annotations-Field prior to assigning the result to the field to 
>>> avoid concurrency-effects on the double-checked locking.
>>>
>>> Since Code using getAnnotations() always could trap in an 
>>> annotations-result not fitting any more to the class because of 
>>> concurrent redefinition we would not need to take care of this (and 
>>> the current code could cause this already inside "getAnnotations()".
>>>
>>> -Alex
>>>
>>> Am 01.11.2012 um 15:56 schrieb Joe Darcy:
>>>
>>>> On 11/1/2012 7:11 AM, Alan Bateman wrote:
>>>>> On 01/11/2012 13:17, Alexander Knöller wrote:
>>>>>> Hi there.
>>>>>>
>>>>>> java.lang.Class.getAnnotations() always enters a 
>>>>>> synchronized-block, slowing down multi core machines that heavily 
>>>>>> make use of Annotations.
>>>>>> (in our Case we use LoadTimeWeaving in the spring-framework 3.1.2)
>>>>>> We are using sun-jdk 6 which has the same performance-bottleneck.
>>>>>> openjdk7 and 8 show no fix so far, although it looks like it 
>>>>>> might be possible to build a kind of double-checked locking?
>>>>>>
>>>>>> Has this issue ever been persued?
>>>>>>
>>>>>> Special Regards
>>>>>> Alex Knöller
>>>>>>
>>>>> If you have a proposal then I suggest bringing it to core-libs-dev 
>>>>> for discussion. In addition to contention there are other issues 
>>>>> that need attention there too, particularly the potential to 
>>>>> deadlock and the overhead per Class when annotations aren't used. 
>>>>> There's definitely some useful work that could be done there.
>
> Hi all,
>
> initAnnotationsIfNecessary() is at least synchronized, so that it 
> always gets the result right in face of concurrent calls to it and 
> class redefinitions performed by VM, but other methods, such as for 
> example:
>
>     private Field[] privateGetDeclaredFields(boolean publicOnly)
>
> ...and many others are not. Therefore a theoretical race exists that 
> could install an old version of fields into cached storage 
> (declaredPublicFields or declaredFields in this case) that was taken 
> before class was redefined and write it over the new version of fields...

I digress. Javadocs say: "The redefinition may change method bodies, the 
constant pool and attributes. The redefinition must not add, remove or 
rename fields or methods, change the signatures of methods, or change 
inheritance. These restrictions maybe be lifted in future versions."

So currently there's no problem since redefinition can only change 
method bodies, so reflecting over old or new version of the class 
returns the same results.

But the synchronization bottleneck of initAnnotationsIfNecessary() could 
be solved this way.

Regrads, Peter

>
> I suggest redesigning the lazy construction / caching by employing 
> versioned containers like the following:
>
> static class VersionedSoftRef<T> extends SoftReference<T> {
>   final int redefinedCount;
>   VersionedSoftRef(T referent,  int redefinedCount) {
>     super(referent);
>     this.redefinedCount = redefinedCount;
>   }
> }
>
> ...to be used instead of plain SoftReferences in places like 
> declaredPublicFields and such and using CAS (via Unsafe or 
> AtomicReferenceFieldUpdater) to optimistically install lazily 
> constructed data...  The versioned containers would serve two 
> purposes: each access to a part of cached data could be independently 
> version-checked against current value of classRedefinedCount on a fast 
> path before returning a cached value. In case of cache-miss (not data 
> or stale data), a thread could compute the data concurrently with any 
> other threads doing the same and using CAS at the end, install the 
> latest version of data into cache.
>
> For getAnnotations() I would use a similar technique (a versioned 
> private subclass of HashMap for example).
>
> If you like, I can prepare a patch and send it for review.
>
>
> Regards, Peter
>
>>>>>
>>>>>
>>>> Note that the block in question is synchronized so that 
>>>> getAnnotations returns the right result if the class has been 
>>>> redefined at runtime using an API for that purpose.
>>>>
>>>> -Joe
>>>>
>