[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH] xen: reset creation_finished flag on soft reset

> -----Original Message-----
> From: Vitaly Kuznetsov [mailto:vkuznets@xxxxxxxxxx]
> Sent: 01 September 2017 10:27
> To: Andrew Cooper <Andrew.Cooper3@xxxxxxxxxx>; Paul Durrant
> <Paul.Durrant@xxxxxxxxxx>
> Cc: xen-devel@xxxxxxxxxxxxx; George Dunlap <George.Dunlap@xxxxxxxxxx>;
> Ian Jackson <Ian.Jackson@xxxxxxxxxx>; Jan Beulich <jbeulich@xxxxxxxx>;
> Konrad Rzeszutek Wilk <konrad.wilk@xxxxxxxxxx>; Stefano Stabellini
> <sstabellini@xxxxxxxxxx>; Tim (Xen.org) <tim@xxxxxxx>; Wei Liu
> <wei.liu2@xxxxxxxxxx>
> Subject: Re: [PATCH] xen: reset creation_finished flag on soft reset
> Andrew Cooper <andrew.cooper3@xxxxxxxxxx> writes:
> > On 01/09/2017 10:11, Vitaly Kuznetsov wrote:
> >> C/s e7dabe5 ("x86/hvm: don't unconditionally create a default ioreq
> >> server") broke soft reset when QEMU traditional is being used. During
> >> soft reset QEMU is relaunched and default ioreq server needs to be
> >> re-created upon first HVM_PARAM_*IOREQ_* request. The flag will be
> >> set back to 'true' when toolstack unpauses the domain, just like after
> >> normal creation.
> >>
> >> Signed-off-by: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>
> >
> > Sorry, but nack.  d->creation_finished is used for a number of things,
> > one being TLB safety before the vcpus have started executing.
> >
> > We either need to split the variable, or rework e7dabe5 to not use this.
> >
> I think that adding another flag is a bad idea, even 'creation_finished'
> flag looks a bit hackish to me. Adjusting e7dabe5 is probably
> better. However, while reading its blurb I don't fully understand the
> change: on migration we create new domain and thus reset
> creation_finished. During QEMU launch we still need to create ioreq
> server. Paul, could you please elaborate a bit (e.g. what are we
> guarding against, when creating ioreq server is redundant) so we can
> suggest a fix for soft reset?

My memory is hazy as to the exact problem, but I think it was an issue with the 
COLO project. IIRC they repeatedly 'migrate' a VM but then resume the original. 
Without e7dabe5 the sending VM ends up with a default ioreq server after the 
first migration because the save code reads the HVM params that trigger its 

I wonder whether the easiest thing to do would be to modify qemu trad to do 
explicit ioreq server creation? It's really not that much code-change... 20-30 
lines or so.


> Thanks,
> --
>   Vitaly
Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.