[jboss-jira] [JBoss JIRA] Commented: (JGRP-532) TCP: hangs without FC

[Bela Ban (JIRA)]

    [ http://jira.jboss.com/jira/browse/JGRP-532?page=comments#action_12365526 ] 
            
Bela Ban commented on JGRP-532:
-------------------------------

So, the basic issue is this:
- We have 1 thread that receives messages from each node (Connection.Receiver thread)
- When all threads are used up in the pool, rejection policy of "run" make that receiver thread deliver the message
- If that's the case and the receiver thread blocks on NAKACK's NakReceiverWindow, then it will not be able to pull further data off of the socket. After the socket's receive window fills up, the sender will block in write(). Alternatively, if use_send_queues is true, the receiver won't take messages off the queue, so put() will block (as the queue is bounded).
- How can the receiver thread block ? If another receiver thread ahead of it acquires the lock for the NakReceiverWindow in NAKACK and calls down() on the same thread and happens to block on write() !
Example:
- Threads 1,2 and 3 have submitted messages to the thread pool
- Now the thread pool is exhausted, so "run" makes sure the Connection.Receiver calls up() directly (on its own thread)
- The Connection.Receiver blocks until it can acquire the NRW lock, therefore it doesn't pull data off of the socket
- Threads 1-3 send data, at one point, the TCP write() might block, say for example for Thread 3.
- Thread-3 now blocks on write() and has the NRW lock
- The Connection.Receiver thread waits for the NRW lock, but meanwhile doesn't cause the write() to unblock as it doesn't pull data off of the socket: ===> DEADLOCK !

SOLUTION: threads 1-3 need to unlock the NRW *before* they send down messages ! This is JIRA 535


> TCP: hangs without FC
> ---------------------
>
>                 Key: JGRP-532
>                 URL: http://jira.jboss.com/jira/browse/JGRP-532
>             Project: JGroups
>          Issue Type: Bug
>            Reporter: Bela Ban
>         Assigned To: Bela Ban
>             Fix For: 2.5
>
>         Attachments: config.txt, tcp.xml, tmp.txt
>
>
> Usually, FC is enabled, but when FC is not present, perf.Test with 2 senders and the attached tcp.xml/config.txt hangs. Stack trace attached too.

-- 
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators: http://jira.jboss.com/jira/secure/Administrators.jspa
-
For more information on JIRA, see: http://www.atlassian.com/software/jira