Packets that are sent to the filter queue are usually
sent onwards straight away, but if they can't be
(e.g. the guest is stopped) they sit in the queue until
something causes them to be sent.  Flush the queue
when teh guest restarts.

TODO: I think in a lot of these cases the packets the
queue system stops being sent, should be allowed to be sent when
the guest is stopped.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Just debug enable/disable etc

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

The packet lists are accessed by:
   a) The main thread for incoming packets on the socket and network
   b) the compare thread
   c) The migration/colo thread (when it calls flush)

Mutex them to stop multiple threads changing the lists at the same
time.

These probably need to turn into RCU lists that would avoid the locks?

TODO: Hmm some deadlock potential between the big lock when taken for
the monitor commands and the packet lists as we read them in info
colo-proxy; anywhere else?

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Based on the technique in the earlier kernel colo-proxy.
This only handles inbound connections;  when it sees
the guest send the syn/ack and then the outside world
respond with the ack, it uses the number in the ack
to figure out the sequence number hte primary was using.
Then we have an offset that gets applied to all packets
until the next checkpoint comes in.
This is also applied to incoming ACKs on these connections
so that the secondary guest sees a consistent view.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

More packet stuff; has ack rework

Thread names limit at about 14 chars, and if they're longer they're
ignored.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

On an incoming connection ignore the sequence numbers
on the syn/ack when doing the compare; the secondary
will fix up all future packets to match the primary
after it receives the 'ack' from the outside world.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Add a checkpoint index before each packet sent over the socket
between the VMs.  This is mostly as a sanity check, but it
allows the primary to discard packets that are from the previous
checkpoint.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Add an info command to dump out the state of the proxy; for debug.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Packets read off the wire are big endian; we don't actually
need them to be the correct endian ness internally, but
getting it right makes debug a lot easier.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

When switching to COLO mode force rmeap of all RAM using the new
host addresses.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Add fine grain stats so we can see where the time goes.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Add finer grained stats to the primary side to see where the time
goes in each checkpoint.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Used the TimedAverage type to hold min/max/avg sizes for colo stats
and present them through info migrate.

Based partially on zhanghailiang's earlier stats patches.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Automatically switch into a passive checkpoint mode when checkpoints are
repeatedly short.  This saves CPU time on the SVM (since it's not running)
and the network traffic and PVM CPU time for the comparison processing.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

If the rate of miscompares is very quick we don't make much forward
progress;  start off by limiting the minimum checkpoint period
by adding a delay after we tell the secondary we're going to want
a checkpoint.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Instead of polling colo_proxy's miscompare variable, use a
condition variable and signal it; using timedwait this means the waiting
is free. (Note we haven't got a wrapper for posix_cond_timedwait
possibly because it's not portable?)

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Emit an event each time we sync the dirty bitmap on the source;
this helps libvirt use postcopy by giving it a kick when it
might be a good idea to start the postcopy.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Amit Shah <amit.shah@redhat.com>
Message-Id: <1450266458-3178-5-git-send-email-dgilbert@redhat.com>
Signed-off-by: Amit Shah <amit.shah@redhat.com>

I missed the calls to send migration events on the destination side
as we enter postcopy.
Take care when adding them not to do it after state has been freed.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: zhanghailiang <zhang.zhanghailiang@huawei.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Reviewed-by: Amit Shah <amit.shah@redhat.com>
Message-Id: <1450266458-3178-4-git-send-email-dgilbert@redhat.com>
Signed-off-by: Amit Shah <amit.shah@redhat.com>

This is part of Li Zhijian's 'integrate colo proxy to colo module'

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Minimal verison of Li Zhijian's 'remove netfilter related code'

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Signed-off-by: Li Zhijian <lizhijian@cn.fujitsu.com>

Signed-off-by: Li Zhijian <lizhijian@cn.fujitsu.com>

Signed-off-by: Li Zhijian <lizhijian@cn.fujitsu.com>
dgilbert: merged

Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>

If quorum has two children(A, B). A do flush sucessfully, but B flush failed.
We MUST choice A as winner rather than just pick anyone of them. Otherwise
the filesystem of guest will become read-only with following errors:

end_request: I/O error, dev vda, sector 11159960
Aborting journal on device vda3-8
EXT4-fs error (device vda3): ext4_journal_start_sb:327: Detected abort journal
EXT4-fs (vda3): Remounting filesystem read-only

Signed-off-by: Wen Congyang <wency@cn.fujitsu.com>
Signed-off-by: Changlong Xie <xiecl.fnst@cn.fujitsu.com>

Commit c8ee0a4 introduced new events containing PRIx64 constants without
including the % prefix in the preceding string. This results in a compile
error during build if --enable-trace-backends is passed to configure.

Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>
Message-id: 1450566522-6003-1-git-send-email-mark.cave-ayland@ilande.co.uk
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>

Implement a 'shutdown' like function.
RDMA doesn't have a direct equivalent of the shutdown call;
so set low timeouts and try and force the connection into an error
state.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

Group write notifications into a single message for multiple
notifications to reduce overhead.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

For debug only; the 'ready' messages sent on the control
path currently contain no data; adding a sequencer counter
helps in debugging.

TODO: Make this safe if receiving a stream from an older
QEMU.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

When RDMA receives write-notify messages, inform the ram code.
This interface also gets called where RDMA places a zero page.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

RDMA has the advantage that data written to the destination
just appears in the destination memory space; the downside of this is
for COLO (and postcopy) where they have to do something to the
received data.  Here they need a notification that pages have been
updated.

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>