[
https://issues.apache.org/jira/browse/ZOOKEEPER-2091?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14244861#comment-14244861
]
Camille Fournier commented on ZOOKEEPER-2091:
---------------------------------------------
I'm all the way in the bottom of this code and don't even remember the
invariants. Why don't we wait till all the bytes are flushed to the buffer if
we don't read anything until that's done?
FWIW, I looked into the depth's of Netty's NIO socket handling. In particular
to verify that Netty sockets will sometimes switch to reading before all of the
outgoing bytes are written, which they will. I thought it was interesting
though that they also configure a "writeSpinCount" that lets them try to flush
all bytes a configurable number of times before going back to the selector to
do other IO. Might be something worth considering. Or, you know, not doing this
socket stuff ourselves at all and just using netty :)
https://github.com/netty/netty/blob/0eb059bf58642f3c06144e1ea4a9d6c7632eb4d5/transport/src/main/java/io/netty/channel/nio/AbstractNioByteChannel.java#L198
In conclusion, I think that it may not matter much either way but probably we
should only add this to the SendPacket method.
> Possible logic error in ClientCnxnSocketNIO
> -------------------------------------------
>
> Key: ZOOKEEPER-2091
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2091
> Project: ZooKeeper
> Issue Type: Bug
> Components: java client
> Affects Versions: 3.4.6
> Reporter: Cheng
> Assignee: Rakesh R
> Fix For: 3.5.1
>
> Attachments: ZOOKEEPER-2091.patch
>
>
> When SASL authentication is enabled, the ZooKeeper client will finally call
> ClientCnxnSocketNIO#sendPacket(Packet p) to send a packet to server:
> @Override
> void sendPacket(Packet p) throws IOException {
> SocketChannel sock = (SocketChannel) sockKey.channel();
> if (sock == null) {
> throw new IOException("Socket is null!");
> }
> p.createBB();
> ByteBuffer pbb = p.bb;
> sock.write(pbb);
> }
> One problem I can see is that the sock is non-blocking, so when the sock's
> output buffer is full(theoretically), only part of the Packet is sent out and
> the communication will break.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
