live migration from legacy to staging fails

live migration from legacy to staging fails

[Olaf Hering]

convert-legacy-stream.py fails to receive a HVM domU coming from xen-4.5.
It runs into legacy.CHUNK_enable_verify_mode "Unable to convert a debug
stream", but only if the '--debug' knob was passed to 'xl migrate'.
The comment there is clearly wrong.

How is this supposed to be handled? I think --debug is supposed to
enable just all the DPRINTKs, nothing else.

Olaf

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel

Re: live migration from legacy to staging fails

[Andrew Cooper]

On 07/04/16 16:40, Olaf Hering wrote:
> convert-legacy-stream.py fails to receive a HVM domU coming from xen-4.5.
> It runs into legacy.CHUNK_enable_verify_mode "Unable to convert a debug
> stream", but only if the '--debug' knob was passed to 'xl migrate'.
> The comment there is clearly wrong.
>
> How is this supposed to be handled? I think --debug is supposed to
> enable just all the DPRINTKs, nothing else.

Sorry - I have been having email problems today.  I see have already
found and addressed the issue.

To answer this question, --debug does more than just increase the
verbosity.  Once the live phase is complete, it sends the verify_mode
marker, then resends the entire memory contents.  (Once the legacy
sending side was complete, it reliably fell over a SIGSEGV for me, and
ended up terminating the migration due to an early eof on the fd.)

The receiving side, on receipt of verify_mode, switches from memcpy() to
memcmp() to verify that the memory it received during the live phase
matches up.

The problematic issue is that any domain with PV rings will expect to
see memory corruption according to verify mode, and this is expected
behaviour due to {net,blk}backends continuing to process requests after
the domain has paused.

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel