Re: [Qemu-devel] [PATCH v4 0/9] enable numa configuration before machine_init() from QMP

[Eduardo Habkost <ehabkost@redhat.com>]

On Mon, Apr 23, 2018 at 06:55:14PM +0200, Igor Mammedov wrote:
> On Mon, 23 Apr 2018 10:05:54 -0300
> Eduardo Habkost <ehabkost@redhat.com> wrote:
> 
> > On Mon, Apr 23, 2018 at 11:50:16AM +0200, Igor Mammedov wrote:
> > > On Fri, 20 Apr 2018 08:31:18 +0200
> > > Markus Armbruster <armbru@redhat.com> wrote:
> > >   
> > > > Eduardo Habkost <ehabkost@redhat.com> writes:
> > > >   
> > > > > On Thu, Apr 19, 2018 at 10:00:04AM +0200, Igor Mammedov wrote:    
> > > > >> On Wed, 18 Apr 2018 09:08:30 +0200
> > > > >> Markus Armbruster <armbru@redhat.com> wrote:
> > > > >>     
> > > > >> > Eduardo Habkost <ehabkost@redhat.com> writes:
> > > > >> >     
> > > > >> > > On Tue, Apr 17, 2018 at 05:41:10PM +0200, Igor Mammedov wrote:      
> > > > >> > >> On Tue, 17 Apr 2018 11:27:39 -0300
> > > > >> > >> Eduardo Habkost <ehabkost@redhat.com> wrote:
> > > > >> > >>       
> > > > >> > >> > On Tue, Apr 17, 2018 at 04:13:34PM +0200, Markus Armbruster wrote:      
> > > > >> > >> > > Igor Mammedov <imammedo@redhat.com> writes:
> > > > >> > >> > > 
> > > > >> > >> > > [...]        
> > > > >> > >> > > > Series allows to configure NUMA mapping at runtime using QMP
> > > > >> > >> > > > interface. For that to happen it introduces a new '-preconfig' CLI option
> > > > >> > >> > > > which allows to pause QEMU before machine_init() is run and
> > > > >> > >> > > > adds new set-numa-node QMP command which in conjunction with
> > > > >> > >> > > > query-hotpluggable-cpus allows to configure NUMA mapping for cpus.
> > > > >> > >> > > >
> > > > >> > >> > > > Later we can modify other commands to run early, for example device_add.
> > > > >> > >> > > > I recall SPAPR had problem when libvirt started QEMU with -S and, while it's
> > > > >> > >> > > > paused, added CPUs with device_add. Intent was to coldplug CPUs (but at that
> > > > >> > >> > > > stage it's considered hotplug already), so SPAPR had to work around the issue.        
> > > > >> > >> > > 
> > > > >> > >> > > That instance is just stupidity / laziness, I think: we consider any
> > > > >> > >> > > plug after machine creation a hot plug.  Real machines remain cold until
> > > > >> > >> > > you press the power button.  Our virtual machines should remain cold
> > > > >> > >> > > until they start running, i.e. with -S until the first "cont".      
> > > > >> > >> It probably would be too risky to change semantics of -S from hotplug to coldplug.
> > > > >> > >> But even if we were easy it won't matter in case if dynamic configuration
> > > > >> > >> done properly. More on it below.
> > > > >> > >>       
> > > > >> > >> > > I vaguely remember me asking this before, but your answer didn't make it
> > > > >> > >> > > into this cover letter, which gives me a pretext to ask again instead of
> > > > >> > >> > > looking it up in the archives: what exactly prevents us from keeping the
> > > > >> > >> > > machine cold enough for numa configuration until the first "cont"?        
> > > > >> > >> > 
> > > > >> > >> > I also think this would be better, but it seems to be difficult
> > > > >> > >> > in practice, see:
> > > > >> > >> > http://mid.mail-archive.com/20180323210532.GD28161@localhost.localdomain      
> > > > >> > >> 
> > > > >> > >> In addition to Eduardo's reply, here is what I've answered back
> > > > >> > >> when you've asked question the 1st time (v2 late at -S pause point reconfig):
> > > > >> > >> https://www.mail-archive.com/qemu-devel@nongnu.org/msg504140.html
> > > > >> > >> 
> > > > >> > >> In short:
> > > > >> > >> I think it's wrong in general doing fixups after machine is build
> > > > >> > >> instead of getting correct configuration before building machine.
> > > > >> > >> That's going to be complex and fragile and might be hard to do at
> > > > >> > >> all depending on what we are fixing up.      
> > > > >> > >
> > > > >> > > What "building the machine" should mean, exactly, for external
> > > > >> > > users?    
> > > > >> under "building machine", I've meant machine_run_board_init()
> > > > >> and all follow up steps to machine_done stage.
> > > > >>     
> > > > >> > > The main question I'd like to see answered is: why exactly we
> > > > >> > > must "build" the machine before the first "cont" is issued when
> > > > >> > > using -S?  Why can't we delay everything to "cont" when using -S?      
> > > > >> Nor sure what question is about,
> > > > >> Did you mean if it were possible to postpone machine_run_board_init()
> > > > >> and all later steps to -S/cont time?    
> > > (1)
> > > As David said -S pause point is practically breakpoint on some
> > > instruction of built/existing machine and current monitor commands
> > > expect it to be valid. Moving -S before machine_run_board_init()
> > > will break semantics of current -S pause point (i.e. user expectation
> > > on existing machine) as well as most of the commands that evolved
> > > in environment where machine already existed.  
> > 
> > OK, so what's missing here is a clear description what the user
> > can expect on -S.
> Currently it's fully configured machine with all CLI options taken
> in account in paused state in initial state or with state it is getting
> from migration stream if -incoming were used in combination with -S.
> 
> > > Hence a new -preconfig option and runstate to avoid breaking
> > > exiting users and being able to cleanly handle configuration that
> > > affects machine_run_board_init().
> > >   
> > > > > Exactly.  In other words, what exactly must be done before the
> > > > > monitor is available when using -S,  
> > > for MUST, it should be commands that affect machine_run_board_init()
> > > like being added set-numa-node
> > >   
> > > > > and what exactly can be postponed after "cont" when using -S?  
> > > hotplug configuration and various runtime query commands that
> > > expect built machine. (today it's most of the commands)
> > > 
> > > wrt configuration commands we should split them into coldplug
> > > and hotplug ones (some could be both).
> > >      
> > > > >> > > Is it just because it's a long and complex task?  Does that mean
> > > > >> > > we might still do that eventually, and eliminate the
> > > > >> > > prelaunch/preconfig distinction in the distant future?      
> > > > >> > 
> > > > >> > Why would anyone want to use -S going forward?  For reasons other "we've
> > > > >> > always used -S, and can't be bothered to change".    
> > > > >> We should be able to deprecate/remove -S once we can do all
> > > > >> initial configuration that's possible to do there at
> > > > >> preconfig time.    
> > > > 
> > > > This sounds like there are things we can do with -S but can't
> > > > --preconfig now.  Is that correct?  
> > > yes, we can't do at --preconfig time anything that requires built machine.  
> > 
> > "built machine" is a very broad description.  We need to specify
> > more clearly what "built machine" means for an external user.
> > Does it mean having the QOM tree available?  Does it mean having
> > the VCPU threads created?  Without defining what -S really must
> > provide, we won't be able to deprecate and replace it.
> (*2) how about s/built machine/machine ready to execute guest code/,
> that's what it is now.

This is a bit better, we still need to be clear about what
"ready" means.  e.g.: can users expect the VCPU threads be
already running?

Anyway, the details don't need to be sorted out immediately.  IMO
the most important part is to describe the difference between
-preconfig and -S.

> 
> 
> > > > > If the plan is to deprecate -S, what are the important
> > > > > user-visible differences between -S and -preconfig today?  Do we
> > > > > plan to eliminate all those differences before
> > > > > deprecating/removing -S?  
> > > we probably won't be able to deprecate -S in foreseeable future,
> > > for that we would need to be able to do everything starting from
> > > machine_run_board_init() to current pause point.
> > > But we can gradually move configuration commands to -preconfig time
> > > and gradually add CLI equivalents for that aren't possible at -S time
> > > (like Paolo suggested picking to be used machine model at runtime)  
> > 
> > This could be a good plan, if we can explain why exactly -S is
> > still needed.
> For a while -S would be need at least for compat reasons, if we ever
> get to point where at -preconfig time machine could be build up to the
> point -S provides[2] then we can talk about deprecating it, for now it's
> way too premature to do something about it /I mean documenting intent
> which is not there yet and might never materialize as there is no real
> demand to deprecate it/.

Yeah, compatibility is the main reason we can't simply deprecate
or remove -S immediately.  We just need to find out what exactly
is important on -S.


> 
> > [...]
> > > > >>                       But I've been sitting on these patches for
> > > > >> a long time and what's obvious to me might be not so clear to others.    
> > > > 
> > > > Par for the course, don't feel bad about it.
> > > >   
> > > > >> I might just not see what's missing. Any suggestions to improve it
> > > > >> are welcome.    
> > > > >
> > > > > I miss something that documents why both -S and -preconfig need
> > > > > to exist, what are the differences between them today, and what
> > > > > we plan to do about the differences between them in the future.  
> > > Where would you prefer it being documented?  
> > 
> > I suggest qemu-options.hx and/or qemu-doc.texi.
> Regarding qemu-options.hx patch
>  "[PATCH for-2.13 v5 03/11] cli: add --preconfig option" 
> adds doc text describing --preconfig option with explanation of how
> 'cont' could be used (including in combination with -S).
> 
> I'll try to come up with a text for qemu-doc.texi, not about
> deprecating -S but about when --preconfig should be used vs -S
> and where to get list of commands that could be used at preconfig state.

Sounds good to me.  Thanks!


> 
> > BTW, "cont" is documented as "Resume guest VCPU execution", which
> > is not true when using preconfig.  Maybe it's better to add a
> > separate QMP command for "create machine and devices" instead of
> > overloading the semantics of "cont"?
> My bad, I've missed it, I can fixup 'cont' description to match
> its behavior with --preconfig taken in account.
> 
> I'm not so sure about adding a new command is better though, I recall
> Markus being against adding new commands unless we have to,
> but I don't have strong inclination both ways so it's up to you.
> 
> I'm more inclined towards reusing 'cont', it seems logical 
> (/me looking from the point if I were user).

'cont' seemed logical to me at first, until I read its
documentation.  Then I think it makes things very confusing,
especially if we combine -preconfig with -S and/or -incoming.

A separate command would have less room for ambiguity.

-- 
Eduardo