Re: [PATCH v7 07/10] vfio: Extend the device migration protocol with PRE_COPY

[Alex Williamson]

On Thu, 3 Mar 2022 19:49:51 -0400
Jason Gunthorpe <jgg@xxxxxxxxxx> wrote:

> On Thu, Mar 03, 2022 at 12:59:30PM -0700, Alex Williamson wrote:
> 
> > > > If it's an abuse, then let's not do it.  It was never my
> > > > impression or intention  
> 
> So maybe abuse is the wrong word, but I don't want to mess up this
> interface, which is intended to support real pre-copy devices, just
> because devices that don't actually implement true precopy might do
> silly things.

Abuse... silly... either way, you're clearly not comfortable misusing
PRE_COPY for this purpose.  

> The vGPU case you imagine will still work and qemu will switch to
> STOP_COPY with a huge trailer and be slow. That is unavoidable and I
> think it is fine.

It's not really fine, but I think it will require some better defined
interfaces and userspace support to give a clear picture of how data is
partitioned.

> > > > Furthermore the acc driver was explicitly directed not to indicate any degree
> > > > of trailing data size in dirty_bytes, so while trailing data may be small for acc,
> > > > this interface is explicitly not intended to provide any indication of trailing
> > > > data size.  Thanks,   
> 
> Yes, trailing data is not what this is for. This is only to help
> decide when to switch from PRE_COPY to STOP_COPY. If the device can
> execute STOP_COPY in the right time is a completely different
> discussion/interface.
> 
> > > Just to clarify, so the suggestion here is not to use PRE_COPY for compatibility
> > > check at all and have a different proper infrastructure for that later as Jason
> > > suggested?
> > > 
> > > If so, I will remove this patch from this series and go back to the old revision
> > > where we only have STOP_COPY and do the compatibility check during the final
> > > load data operation.  
> > 
> > Hi Shameer,
> > 
> > I think NVIDIA has a company long weekend, so I'm not sure how quickly
> > we'll hear a rebuttal from Jason, but at this point I'd rather not
> > move  
> 
> Yes, company long weekend.
> 
> > forward with using PRE_COPY exclusively for compatibility testing if
> > that is seen as an abuse of the interface, regardless of the size of
> > the remaining STOP_COPY data.  It might be most expedient to respin
> > without PRE_COPY and we'll revisit methods to perform early
> > compatibility testing in the future.  Thanks,  
> 
> Shameerali has talked about wanting this compat check early from the
> start, and done all the work to implement it. I think it is pretty
> extreme to blow up his series over trailing_data.
> 
> To me acc is fine to use it this way until we get a better solution
> for compatability. We all need this, but I expect it to be complicated
> to define.

It was only in v7 that we made this switch to use PRE_COPY for this
purpose, I wouldn't call it blowing up his series to step back and
decide that was a poor choice and clearly v8 exists without this.  This
isn't the end of the discussion regarding early compatibility testing,
but I'm not going to rush a PRE_COPY interface to support that early
compatibility testing if we're not agreed that it's a valid use case,
and not just a marginally acceptable one due to the trailing data being
inconsequential.  Let's focus on v8 and we can talk about further
extensions later.  Thanks,

Alex