Hi folks, On Thu, 2006-12-14 at 11:01 +0100, Daniel Lezcano wrote: > Perhaps, it could work, perhaps not. To use the zero window we should > check that all TCP stacks are compatible with that. If we have a server > running on linux and we checkpoint it, it will not be cool if all > windows client lose their connecions while all linux client or mac > clients stay blocked. I understand what you mean. We should take a conservative way, even if dropping packets results in shrinking the window size. > > It seems we should check there is no packet in receive_queue > > before closing the socket. If there is, the kernel tries to > > send a RESET packet. > > The receive and the send queue are checkpointed. If the receive queue is > not flushed before destroying the socket, there is no issue because the > RST packet will be dropped because of the blocked traffic. You are right. I forgot that all packets would be dropped after checkpointing in MCR. Thanks, Masahiko.