Re: [PATCH V2] migration: file URI

[Peter Xu]

On Wed, Jun 14, 2023 at 12:47:41PM -0300, Fabiano Rosas wrote:
> Peter Xu <peterx@redhat.com> writes:
> 
> > On Mon, Jun 12, 2023 at 03:39:34PM -0400, Steven Sistare wrote:
> >> On 6/12/2023 2:44 PM, Peter Xu wrote:
> >> > Hi, Steve,
> >> > 
> >> > On Wed, Jun 07, 2023 at 11:38:59AM -0700, Steve Sistare wrote:
> >> >> Extend the migration URI to support file:<filename>.  This can be used 
> >> >> for
> >> >> any migration scenario that does not require a reverse path.  It can be 
> >> >> used
> >> >> as an alternative to 'exec:cat > file' in minimized containers that do 
> >> >> not
> >> >> contain /bin/sh, and it is easier to use than the fd:<fdname> URI.  It 
> >> >> can
> >> >> be used in HMP commands, and as a qemu command-line parameter.
> >> > 
> >> > I have similar question on the fixed-ram work,
> >> 
> >> Sorry, what is the "fixed-ram work"?
> >
> > https://lore.kernel.org/all/20230330180336.2791-1-farosas@suse.de
> >
> > It has similar requirement to migrate to a file, though slightly different
> > use case.
> >
> >> 
> >> > on whether we should assume
> >> > the vm stopped before doing so.  Again, it leaves us space for
> >> > optimizations on top without breaking anyone.
> >> 
> >> I do not assume the vm is stopped.  The migration code will stop the vm
> >> in migration_iteration_finish.
> >> 
> >> > The other thing is considering a very busy guest, migration may not even
> >> > converge for "file:" URI (the same to other URIs) but I think that 
> >> > doesn't
> >> > make much sense to not converge for a "file:" URI.  The user might be 
> >> > very
> >> > confused too.
> >> 
> >> The file URI is mainly intended for the case where guest ram is backed by 
> >> shared memory 
> >> and preserved in place, in which case writes are not tracked and 
> >> convergence is not an
> >> issue.  If not shared memory, the user should be advised to stop the 
> >> machine first.
> >> I should document these notes in qemu-options and/or migration.json.
> >
> > My question was whether we should treat "file:" differently from most of
> > other URIs.  It makes the URI slightly tricky for sure, but it also does
> > make sense to me because "file:" implies more than the rest URIs, where
> > we're sure about the consequence of the migration (vm stops), in that case
> > keeping vm live makes it less performant, and also weird.
> >
> > It doesn't need to be special in memory type being shared, e.g. what if
> > there's a device that contains a lot of data to migrate in the future?
> > Assuming "shared memory will always migrate very fast" may not hold true.
> >
> > Do you think it makes more sense to just always stop VM when migrating to
> > file URI?  Then if someone tries to restart the VM or cancel the migration,
> > we always do both (cancel migration, then start VM).
> 
> From our discussions in the other thread, I have implemented a
> MIGRATION_CAPABILITY_SUSPEND to allow the management layer to decide
> whether the guest should be stopped by QEMU before the migration.
> 
> I'm not opposed to coupling file URI with a stopped VM, although I
> think, at least for fixed-ram, libvirt would prefer to be able to decide
> when to stop.

IIUC the best timing is when migration starts, not earlier, not later.

If that's always the case, it's better qemu guarantee that?  Or am I wrong
that libvirt wants to not do it in some cases?

-- 
Peter Xu