[PING] JDK 10 RFR of 8184157: (ch) AsynchronousFileChannel hangs with internal error when reading locked file

Roger Riggs Roger.Riggs at Oracle.com
Fri Aug 11 13:49:55 UTC 2017


Hi Brian,

I'm not sure I have enough background to fully review this.

Seeing that overlapped (some) memory is freed only at close, can there 
be a problem with
a very active channel that is open for a long time to accumulate/leak 
native memory?
PendingIoCache:135-136: Or is that just belt and suspenders?

Thanks, Roger


On 8/8/2017 4:11 PM, Brian Burkhalter wrote:
> Begin forwarded message:
>
>> *From: *Brian Burkhalter <brian.burkhalter at oracle.com 
>> <mailto:brian.burkhalter at oracle.com>>
>> *Subject: **JDK 10 RFR of 8184157: (ch) AsynchronousFileChannel hangs 
>> with internal error when reading locked file*
>> *Date: *July 24, 2017 at 3:54:55 PM PDT
>> *To: *nio-dev <nio-dev at openjdk.java.net 
>> <mailto:nio-dev at openjdk.java.net>>
>>
>> https://bugs.openjdk.java.net/browse/JDK-8184157
>> http://cr.openjdk.java.net/~bpb/8184157/webrev.00/
>>
>> This problem is caused by a race condition wherein the PendingFuture 
>> of a lock operation in the PendingIoCache is replaced by that of a 
>> read operation, with both using the same key, before the completion 
>> status of the lock operation is dequeued by the event handler thread 
>> of Iocp. When the lock completion status is dequeued, the event 
>> handler looks up its corresponding PendingFuture but retrieves that 
>> of the read operation instead. This causes the completed() callback 
>> of the read task to be invoked with the NumberOfBytes transferred for 
>> the lock operation which is a garbage value. This results in an 
>> attempt to set the read ByteBuffer position to a value outside its 
>> permitted range.
>>
>> The proposed solution is to introduce an invalidate() method on the 
>> PendingIoCache which removes the entry corresponding to the 
>> OVERLAPPED pointer from the map of overlapped-to-PendingFuture 
>> entries and places the OVERLAPPED pointer value in a Set where it 
>> remains until remove() is called. This ensures that the OVERLAPPED 
>> pointer map key will not be used for a PendingFuture to which it does 
>> not correspond. The pointer values of invalidated PendingFutures are 
>> removed from the Set when remove() is called in the Iocp event 
>> handler after the corresponding completion status is dequeued. All 
>> existing regression tests pass with this patch applied.
>>
>> Thanks,
>>
>> Brian



More information about the nio-dev mailing list