Re: [PATCH v3 1/1] vhost-user-fs: add migration type property

[Michael S. Tsirkin]

On Wed, Mar 01, 2023 at 05:07:28PM +0200, Anton Kuchin wrote:
> On 28/02/2023 23:24, Michael S. Tsirkin wrote:
> > On Tue, Feb 28, 2023 at 07:59:54PM +0200, Anton Kuchin wrote:
> > > On 28/02/2023 16:57, Michael S. Tsirkin wrote:
> > > > On Tue, Feb 28, 2023 at 04:30:36PM +0200, Anton Kuchin wrote:
> > > > > I really don't understand why and what do you want to check on
> > > > > destination.
> > > > Yes I understand your patch controls source. Let me try to rephrase
> > > > why I think it's better on destination.
> > > > Here's my understanding
> > > > - With vhost-user-fs state lives inside an external daemon.
> > > > A- If after load you connect to the same daemon you can get migration 
> > > > mostly
> > > >     for free.
> > > > B- If you connect to a different daemon then that daemon will need
> > > >     to pass information from original one.
> > > > 
> > > > Is this a fair summary?
> > > > 
> > > > Current solution is to set flag on the source meaning "I have an
> > > > orchestration tool that will make sure that either A or B is correct".
> > > > 
> > > > However both A and B can only be known when destination is known.
> > > > Especially as long as what we are really trying to do is just allow qemu
> > > > restarts, Checking the flag on load will thus achive it in a cleaner
> > > > way, in that orchestration tool can reasonably keep the flag
> > > > clear normally and only set it if restarting qemu locally.
> > > > 
> > > > 
> > > > By comparison, with your approach orchestration tool will have
> > > > to either always set the flag (risky since then we lose the
> > > > extra check that we coded) or keep it clear and set before migration
> > > > (complex).
> > > > 
> > > > I hope I explained what and why I want to check.
> > > > 
> > > > I am far from a vhost-user-fs expert so maybe I am wrong but
> > > > I wanted to make sure I got the point across even if other
> > > > disagree.
> > > > 
> > > Thank you for the explanation. Now I understand your concerns.
> > > 
> > > You are right about this mechanism being a bit risky if orchestrator is
> > > not using it properly or clunky if it is used in a safest possible way.
> > > That's why first attempt of this feature was with migration capability
> > > to let orchestrator choose behavior right at the moment of migration.
> > > But it has its own problems.
> > > 
> > > We can't move this check only to destination because one of main goals
> > > was to prevent orchestrators that are unaware of vhost-user-fs specifics
> > > from accidentally migrating such VMs. We can't rely here entirely on
> > > destination to block this because if VM is migrated to file and then
> > > can't be loaded by destination there is no way to fallback and resume
> > > the source so we need to have some kind of blocker on source by default.
> > Interesting.  Why is there no way? Just load it back on source? Isn't
> > this how any other load failure is managed? Because for sure you
> > need to manage these, they will happen.
> 
> Because source can be already terminated

So start it again.

> and if load is not supported by
> orchestrator and backend stream can't be loaded on source too.

How can an orchestrator not support load but support migration?

> So we need to
> ensure that only orchestrators that know what they are doing explicitly
> enable
> the feature are allowed to start migration.

that seems par for the course - if you want to use a feature you better
have an idea about how to do it.

If orchestrator is doing things like migrating to file
then scp that file, then it better be prepared to
restart VM on source because sometimes it will fail
on destination.

And an orchestrator that is not clever enough to do it, then it
just should not come up with funky ways to do migration.


> > 
> > > Said that checking on destination would need another flag and the safe
> > > way of using this feature would require managing two flags instead of one
> > > making it even more fragile. So I'd prefer not to make it more complex.
> > > 
> > > In my opinion the best way to use this property by orchestrator is to
> > > leave default unmigratable behavior at start and just before migration 
> > > when
> > > destination is known enumerate all vhost-user-fs devices and set 
> > > properties
> > > according to their backends capability with QMP like you mentioned. This
> > > gives us single point of making the decision for each device and avoids
> > > guessing future at VM start.
> > this means that you need to remember what the values were and then
> > any failure on destination requires you to go back and set them
> > to original values. With possibility of crashes on the orchestrator
> > you also need to recall the temporary values in some file ...
> > This is huge complexity much worse than two flags.
> > 
> > Assuming we need two let's see whether just reload on source is good
> > enough.
> 
> Reload on source can't be guaranteed to work too. And even if we could
> guarantee it to work then we would also need to setup its incoming migration
> type in case outgoing migration fails.

Since it's local you naturally just set it to allow load. It's trivial - just
a command line property no games with QOM and no state.


> If orchestrator crashes and restarts it can revert flags for all devices

revert to what?

> or can rely on next migration code to setup them correctly because they have
> no effect between migrations anyway.

but the whole reason we have this stuff is to protect against
an orchestrator that forgets to do it.

> Reverting migration that failed on destination is not an easy task too.
> It seems to be much more complicated than refusing to migrate on source.

It is only more complicated because you do not consider that
migration can fail even if QEMU allows it.

Imagine that you start playing with features through QOM.
Now you start migration, it fails for some reason (e.g. a network
issue), and you are left with a misconfigured feature.

Your answer is basically that we don't need this protection at all,
we can trust orchestrators to do the right thing.
In that case just drop the blocker and be done with it.


> I believe we should perform sanity checks if we have data but engineering
> additional checks and putting extra restrictions just to prevent
> orchestrator
> from doing wrong things is an overkill.

Exactly. The check on source is such an overkill - your problem
is not on source, source has no issue sending the VM. Your problem is
on destination - it can not get the data from daemon since the daemon
is not local.


> > > But allowing setup via command-line is valid too because some backends may
> > > always be capable of external migration independent of hosts and don't 
> > > need
> > > the manipulations with QMP before migration at all.
> > I am much more worried that the realistic schenario is hard to manage
> > safely than about theoretical state migrating backends that don't exist.
> > 
> >