On 04/06/2016 09:52 PM, Amit Shah wrote: > On (Wed) 06 Apr 2016 [09:48:19], Jason Wang wrote: >> >> On 04/05/2016 09:32 PM, Dr. David Alan Gilbert wrote: >>> * Amit Shah (amit.s...@redhat.com) wrote: >>>> On (Tue) 23 Feb 2016 [15:02:58], Jason Wang wrote: >>>>>>> This means that 2.5 cannot migrate 2.4 virtual machines, right? Is that >>>>>>> something we want to rectify in 2.6 by making e1000-82540em an alias of >>>>>>> e1000 (instead of the other way round)? >>>>>> You're right; I misread it. With that commit (8304402033): >>>>>> >>>>>> 2.4 with e1000-82540em will not migrate to 2.5 with e1000-82540em. >>>>>> >>>>>> This is despite they're aliased (so the cmdline is backward >>>>>> compatible), but the migration device name actually changed. >>>>>> >>>>>> Of course, 2.5->2.4 will also not work. >>>>>> >>>>>> Since 2.4 emits 'e1000-82540em' as the device name in the migration >>>>>> stream, and 2.5 emits just 'e1000', we have two different names for >>>>>> the same device in two versions. >>>>>> >>>>>> To fix this, we'll need a hack on the dest side to allow e1000 and >>>>>> e1000-82540em in the migration stream for the device, and this can be >>>>>> done for 2.6 and 2.5.stable. >>>>>> >>>>>> Jason, can you attempt this? >>>>>> >>>>> Sure, but just need to understand the "problem". If I understand this >>>>> correctly, the issue only happen for JSON description at the end of >>>>> migration stream, and it won't break migration in fact? >>>> No, this does break migration. >>>> >>>> The stream contained 'e1000-82540em' as the section header for the >>>> device earlier. It now only has 'e1000'. So a newer qemu will only >>>> accept 'e1000', but not 'e1000-82540em' (from an older release). >>> OK, so do we need to get this fixed for 2.6 - i.e. now. >>> >>> Dave >> Sorry for the late response. >> >> Have a try with 2.4 -> 2.5 migration and it works. Looking at >> save_section_header(), it will save se->idstr which seems always be >> "e1000" even if e1000-82540em is used in cli, or is there anything I missed? > OK, sorry for the wrong alarm. > > The VMStateDescription's 'name' field has remained the same; but the > section name has changed, which is fine (that isn't transmitted over > the wire). So my initial reading was correct - and the whitelist I > added in 1483e0d74dcfd183ff46dd63cc57e1fe8b775bf8 is fine. > > So nothing needed for the migration stream. > > Thanks, Jason. > > Amit >
Thanks for the confirmation.
