Difference between revisions of "TCP repair TODO"
Jump to navigation
Jump to search
Line 25: | Line 25: | ||
[[Category:Plans]] | [[Category:Plans]] | ||
[[Category:Network]] | [[Category:Network]] | ||
+ | [[Category:Fly in the ointment]] |
Revision as of 08:46, 17 November 2016
Main article: TCP connection
TCP repair feature in the Linux kernel is supposed to help migrating a TCP socket. It's not yet complete and this page lists what is to be done.
- Transitional states
- Currently we support sockets in closed and establised states. However, if a socket is in e.g. syn-sent state the process of turning it into established can last long. We should teach the kernel and criu to checkpoint and restore this and other states.
The half-opened connections on the server side can be dropped dump-time with the --skip-in-flight
option. This helps in heavily loaded cases when such sockets appear often.
- Optimized restore
- Currently the whole outgoing queue is restored in a "all was sent, waiting for ACK-s" state. After this the data that was really not sent yet will be re-transmitted after a while. This will make the connection work, but will delay it for some time. Need to improve this.
- OOB data
- Nothing to say here actually. This data is just not supported currently.
- Window restore fix
- Currently what we do it send the window probe skb when repair is OFF. The other side should send us the response, but this process is not guaranteed to work. Need to fix this either by saving and restoring the window value, or by re-transmitting the probe again and again.
- Shutdown sockets repair
- Need to place checks in the inet shutdown code similar to those on connect/sendmsg paths.
- Connection tracking
- The nf_conntrack thing in the kernel is about to be live-migrated too. There's currently an API for getting the conntrack info (/proc file), but no such for restoring it by hands.