[tipc-discussion] Removal of global node overload limit

[Erik H. <eri...@er...>]

The global overload limit in TIPC need to be removed.

Currently, this is defined to 5000 messages, and
Two limits are enforced based on this constraint, one global on node 
level and one per socket.
The socket level limit is unnecessary since each socket already have a
hard restriction for how much data can be queued based on receive buffer
fill level and the port-level flow control will prevent a fast sender
from overrunning a slow receiver.

The node level limit does not serve any functional purpose, but rather
creates a bug on on nodes that spawns a lot of server sockets, typically
a listen server in a larger cluster. The global node overload protection
will cause unexpected and unnecessary connection drops when messages are
rejected, even if the receiving socket have space to hold the packet.

I hope you agree with me that this is serious since it renders TIPC 
unusable for "heavy users".
I plan to post a patch that removes this if the weekend soak-testing 
goes well.

//E

PS.
Another reason to remove it is the TIPC virtual device feature that i 
have been fiddling around with. It cannot function properly with this in 
place since it pulls skb's directly from the sockets  sk_receive_queue, 
modifies them and passes it on to the device layer.