RFR 8017540: Improve multi-threaded contention behavior of BigInteger.toString() radix conversion cache

[Peter Levart]

On 06/25/2013 01:53 AM, Brian Burkhalter wrote:
> Branching off from this thread
>
> http://mail.openjdk.java.net/pipermail/core-libs-dev/2013-June/018244.html
>
> I filed this issue (should be public tomorrow)
>
> http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=8017540
>
> for the getRadixConversionCache() enhancement. The patch is here
>
> http://cr.openjdk.java.net/~bpb/8017540/
>
> I reviewed and tested the code for correctness and see no issues, so to speak. This patch cannot be integrated prior to that for 4641897.
>
> Thanks,
>
> Brian

Hi,

I think this has similar problem (data-race followed by read from 
half-constructed array) as the prototype that Aleksey constructed here 
(sorry, I have not seen the forked thread then yet):

http://mail.openjdk.java.net/pipermail/core-libs-dev/2013-June/018368.html


     private static BigInteger getRadixConversionCache(int radix, int exponent) {
         BigInteger retVal = null;
         BigInteger[][] pc = powerCache; // volatile read
         BigInteger[] cacheLine = pc[radix];
         int oldSize = cacheLine.length;
         if (exponent >= oldSize) {
             cacheLine = Arrays.copyOf(cacheLine, exponent + 1);
             for (int i = oldSize; i <= exponent; i++) {
                 retVal = cacheLine[i - 1].square();
                 cacheLine[i] = retVal;
             }
             pc[radix] = cacheLine;
             powerCache = pc; // publish by volatile write
         } else {
             retVal = cacheLine[exponent];
         }

         return retVal;
      }



Thread A:
     pc = powerCache;

Thread B:
     pc = powerCache;
     cacheLine = pc[radix];
     oldSize = cacheLine.length;
     // enter if, since exponent >= oldSize
     cacheLine = copyOf(cacheLine, exponent+1);
     // ...for loop filling cacheLine...
     pc[radix] = cacheLine;

Thread A:
     cacheLine = pc[radix]; // this can read new cache line constructed 
by Thread B (data-race)
     oldSize = cacheLine.length;
     // skip if, since exponent < oldSize, enter else
     retVal = cacheLine[exponent]; // this can read null, since 
reference to cacheLine was obtained via data-race
     return retval;


What do you think of the code proposed here:

http://mail.openjdk.java.net/pipermail/core-libs-dev/2013-June/018368.html


Regards, Peter