Re: [PATCH v4 0/8] Draft, incomplete series introducing state migration

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 2236 bytes --]

On Tue, Feb 04, 2025 at 01:47:37AM +0100, Stefano Brivio wrote:
> I applied what I could, and squashed a number patches, including
> those from "[PATCH 0/6] More migration improvements".
> 
> I didn't test the full flow here.

With this patch set, I'm still unable to migrate a connection.  First
there are some simple things:

 * There are, again, some bogus revert hunks in the series which I had
   to remove.  See comments
 * There are some bugs in the already committed patches, I posted a
   couple of fixes for those

Then debugging the actual migration problem.

The first issue I encountered was that the connect() on the target was
failing with EADDRNOTAVAIL.  I think this is a timing issue - the exit
of the source instance isn't happening quite in time to free up the
port.  I was able to work around that by closing the source side
socket immediately after extracting its info with tcp_repair, instead
of postponing that to the exit check_device_state.

To address this properly, I think we need to do a couple of things:

 * Treat the source-side point of no return as being at the end of
   migrate_source().  Close our sockets and purge flows at that point.
 * On the target side, defer actually re-opening/repairing the sockets
   until check_state, or even send_rarp time.

Laurent, if you have some insight (or can find out) what guarantees we
have in terms of what points on the target come after what points on
the source that would be useful to check.


With the hacky workaround, the connect() now success, but I still
couldn't send data from the target.  From strace, the target side
passt doesn't seem to be attempting doing any writes to the socket,
but I'm not sure why yet.  pcap on the target side does show some
packets on the stream coming from the guest, although it's only one
byte (0x0a) and I thought I'd typed 5ish.  It then appears to
retransmit several times without getting an ack from passt.

I'll keep debugging tomorrow if you don't have an insight.
 
-- 
David Gibson (he or they)	| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you, not the other way
				| around.
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]