This page describes how we handle established TCP connections.
TCP repair mode in kernel
TCP_REPAIR socket option was added to the kernel 3.5 to help with C/R for TCP sockets.
When this option is used, a socket is switched into a special mode, in which any action performed on it does not result in anything defined by an appropriate protocol actions, but rather directly puts the socket into a state, in which the socket is expected to be at the end of the successfully finished operation.
For example, calling
connect() on a repaired socket just changes its state to
with the peer address set as requested.
bind() call forcibly binds the socket to a given address (ignoring any potential conflicts).
close() call closes the socket without any transient
socket is silently killed.
To restore the connection properly, bind() and connect() is not enough. One also needs to restore the
TCP sequence numbers. To do so, the
TCP_QUEUE_SEQ options were introduced.
The former one selects which queue (input or output) will be repaired and the latter gets/sets the sequence. Note setting the sequence is only possible on CLOSE-d socket.
Packets in queue
When set the queue to repair as described above, one can call recv or send syscalls on a repaired socket. Both calls
result on peeking or poking data from/to the respective queue. This sounds funny, but yes, for repaired socket one
can receve the outgoing and send the incoming queues. Using the
MSG_PEEK flag for
recv() is required.
There are 4 options that are negotiated by the socket at the connecting stage. These are
- mss_clamp -- the maximum size of the segment peer is ready to accept
- snd _scale -- the scale factor for a window
- sack -- whether selective acks are permitted or not
- tstamp -- whether timestamps on packets are supported
All four can be read with
getsockopt() calls to a socket and in order to restore them the
TCP_REPAIR_OPTIONS sockoption is introduced.
"The sender's timestamp clock is used as a source of monotonic non-decreasing values to stamp the segments"(rfc7323). The Linux kernel uses the jiffies counter as the tcp timestamp.
#define tcp_time_stamp ((__u32)(jiffies))
We add the TCP_TIMESTAMP options to be able to compensate a difference between jiffies counters, when a connection is migrated on another host. When a connection is dumped, criu calls getsockopt(TCP_TIMESTAMP) to get a current timestamp, then on restore it calls setsockopt(TCP_TIMESTAMP) to set this timestamp as a starting point.
Checkpoint and restore TCP connection
With the above sockoptions dumping and restoring TCP connection becomes possible. The criu just reads the socket state and restores it back letting the protocol resurrect the data sequence.
One thing to note here — while the socket is closed between dump and restore the connection should be "locked", i.e. no packets from peer should enter the stack, otherwise the RST will be sent by a kernel. In order to do so a simple netfilter rule is configured that drops all the packets from peer to a socket we're dealing with. This rule sits in the host netfilter tables after the criu dump command finishes and it should be there when you issue the criu restore one.
Another thing to note is -- on restore there should be available the IP address, that was used by the connection. This is automatically so if restore happens on the same box as dump. In case of hand-made live migration the IP address should be copied too.
That said, the command line option
--tcp-established should be used when calling criu to explicitly state, that the
caller is aware of this "transitional" state of the netfilter.
In case the target process lives in NET namespace the connection locking happens the other way. Instead of per-connection iptables rules the "network-lock"/"network-unlock" action scripts are called so that the user could isolate the whole netns from network. Typically this is done by downing the respective veth pair end.