TestBackwardCompat.testCompat410 often fails due to io.netty.util.IllegalReferenceCountException

[apache]

TestBackwardCompat.testCompat410 often fails due to io.netty.util.IllegalReferenceCountException

[eolivelli]

@merlimat @kishorekasi
do you have time to check ? at the moment you are the most netty 4 experts in the group

[kishorekasi]

Enrico, Let me look into this. Kishore

…

On Wed, Jun 21, 2017 at 7:21 AM, Enrico Olivelli ***@***.***> wrote: @merlimat <https://github.com/merlimat> @kishorekasi <https://github.com/kishorekasi> do you have time to check ? at the moment you are the most netty 4 experts in the group — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#198 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADKY-KPG59ic6Mol20bYlMNJ4cBxNNgdks5sGSbagaJpZM4OBCCA> .

[kishorekasi]

Enrico, Found a few refcount leaks, fixed them and created a pull request. I couldn't reproduce the test failure locally. So, could you please check and let me know if this fixes the issue. Kishore

…

On Thu, Jun 22, 2017 at 10:36 AM, Kishore Kasi ***@***.***> wrote: Enrico, Let me look into this. Kishore On Wed, Jun 21, 2017 at 7:21 AM, Enrico Olivelli ***@***.*** > wrote: > @merlimat <https://github.com/merlimat> @kishorekasi > <https://github.com/kishorekasi> > do you have time to check ? at the moment you are the most netty 4 > experts in the group > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub > <#198 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/ADKY-KPG59ic6Mol20bYlMNJ4cBxNNgdks5sGSbagaJpZM4OBCCA> > . >

[sijie]

marked this as a blocker for 4.5.0, since it seems to indicate some kind of reference leaking in current code base.

[sijie]

@eolivelli @kishorekasi

@merlimat and me were looking together at this issue. @merlimat found the problem on how do we handling the invalid op code in v2 protocol.

in TestBackwardCompat#testCompact400, the test is to verify the current client can't talk to a 4.0.0 server.

The 4.5.0 client is sending a protobuf encoded request to 4.0.0 server. The 4.0.0 server will interpret the 4.5.0 protobuf encoded request, but it will realize this is bad request and sending v2 protocol encoded response. because the request is a bad request, 4.0.0 server sent a response back with unknown op code.

In current v2 ResponseEnDecoder (listed as below), when it doesn't know the op code, it will return the buffer to the channel. this might cause the misbehavior in the channel pipeline to decrement reference count.

`
@OverRide
public Object decode(ByteBuf buffer)
throws Exception {
int rc;
long ledgerId, entryId;

        int packetHeader = buffer.readInt();
        byte version = PacketHeader.getVersion(packetHeader);
        byte opCode = PacketHeader.getOpCode(packetHeader);

        switch (opCode) {
        case BookieProtocol.ADDENTRY:
            rc = buffer.readInt();
            ledgerId = buffer.readLong();
            entryId = buffer.readLong();
            return new BookieProtocol.AddResponse(version, rc, ledgerId, entryId);
        case BookieProtocol.READENTRY:
            rc = buffer.readInt();
            ledgerId = buffer.readLong();
            entryId = buffer.readLong();

            if (rc == BookieProtocol.EOK) {
                ByteBuf content = buffer.slice();
                return new BookieProtocol.ReadResponse(version, rc, ledgerId, entryId, content.retain());
            } else {
                return new BookieProtocol.ReadResponse(version, rc, ledgerId, entryId);
            }
        case BookieProtocol.AUTH:
            ByteBufInputStream bufStream = new ByteBufInputStream(buffer);
            BookkeeperProtocol.AuthMessage.Builder builder
                = BookkeeperProtocol.AuthMessage.newBuilder();
            builder.mergeFrom(bufStream, extensionRegistry);
            BookkeeperProtocol.AuthMessage am = builder.build();
            return new BookieProtocol.AuthResponse(version, am);
        default:
            return buffer;
        }
    }

`

One suggested fix from @merlimat is to throw exception in the EnDeCoder when receiving unknown op code. so the netty can close the connection, error out the pending requests and cleaning up the resources.

[eolivelli]

@kishorekasi my error report was against an old version of the master.
I think that @merlimat suggestion is good. Can you give it a try?