Re: [Qemu-devel] [PATCH 10/20] target-i386: prepare cpu_x86_parse_featurestr() to return a set of key, value property pairs

[Igor Mammedov]

On Thu, 20 Dec 2012 20:13:19 -0200
Eduardo Habkost <address@hidden> wrote:

> On Thu, Dec 20, 2012 at 09:22:49PM +0100, Igor Mammedov wrote:
> > On Thu, 20 Dec 2012 12:10:25 -0200
> > Eduardo Habkost <address@hidden> wrote:
> > 
> > It's a log list of answers. please look through them all to the end.
> > > On Wed, Dec 19, 2012 at 09:18:09PM +0100, Igor Mammedov wrote:
> > > > On Wed, 19 Dec 2012 14:54:30 -0200
> > > > Eduardo Habkost <address@hidden> wrote:
> > > > 
> > > > > On Mon, Dec 17, 2012 at 05:01:22PM +0100, Igor Mammedov wrote:
> > > > > > It prepares for converting "+feature,-feature,feature=foo,feature"
> > > > > > into a set of key,value property pairs that will be applied to CPU 
> > > > > > by
> > > > > > cpu_x86_set_props().
> > > > > > 
> > > > > > Each feature handled by cpu_x86_parse_featurestr() will be converted
> > > > > > into foo,val pair and a corresponding property setter by following
> > > > > > patches.
> > > > > > 
> > > > > > Signed-off-by: Igor Mammedov <address@hidden>
> > > > > 
> > > > > Isn't it much simpler to make cpu_x86_parse_featurestr() get the CPU
> > > > > object as parameter and set the properties directly? Or you can see 
> > > > > some
> > > > > use case where saving the property-setting dictionary for later use 
> > > > > will
> > > > > be useful?
> > > > 
> > > > I plan to use cpu_x86_parse_featurestr() + list of properties later when
> > > > we have properties and subclasses in place. Then it would be possible to
> > > > transform cpu_x86_parse_featurestr() + cpu_x86_set_props() into set of
> > > > global properties.
> > > 
> > > Is it really worth it to use global properties when you can simply set
> > > the properties at the code that handles the "-cpu" string (and already
> > > has to create the CPU object)?
> > > 
> > > (more comments about globals are below)
> > > 
> > > 
> > > > 
> > > > In the end the option handler for -cpu XXX,... could be changed into:
> > > > 
> > > > cpu_opt_parser() {
> > > >     // hides legacy target specific ugliness 
> > > >     target_XXX_cpu_compat_parser_callback() {
> > > >         cpu_class_name = get_class_name(optval)
> > > > 
> > > >         // dumb compatibility parser, which converts old format into
> > > >         // canonical form feature,val property list
> > > >         prop_list = parse_featurestr(optval)
> > > 
> > > I'm expecting it to look different: instead of a target-specific parsing
> > > function, you just need target-specific legacy properties on the CPU
> > > object, that would be set by the (generic) featurestr parsing code if
> > > present.
> > > 
> > > (e.g. "legacy-xlevel" could be used for the legacy "1 becomes
> > > 0x80000001" behavior, and "xlevel" would be the more sane xlevel
> > > property without any magic).
> > Introducing 'legacy-xlevel' doesn't solve problem, because with new name
> > compatibility will be broken and user has to fix qemu invocation anyway.
> > The same effect could be achieved if qemu exits with invalid value error and
> > forces user to fix value on qemu invocation without introducing new
> > properties.
> > 
> > An idea of creating legacy properties would be a negative for following
> > reasons:
> >  - new legacy properties with new names could break current users, and force
> >    them to fix their code
> 
> No, the point is to keep compatibility (just like your per-target
> parse_featurestr() approach), the only difference is that compatibility
> details would be encoded on the legacy properties, not inside
> parse_featurestr().
> 
> >  + converting old feature names into the same property names with same
> >    behavior without adding new ones. But sometimes old behavior is wrong or
> >    hard to convert into property, so not good for cpu model eventually.
> 
> That's why the legacy properties would be "legacy" one. The only code
> that would be supposed to use it would be the (legacy) "-cpu" parsing
> code.
> 
> >  - it burdens CPU object with properties that do almost the same as its
> >    'correct' counterpart but not quite so. That complicates API of CPU
> >    object a lot.
> 
> I believe the legacy property is as heavy a burden as a legacy
> featurestr-parsing code. It's no burden to the CPU object to have
> property setters that are almost never used. They would be just sitting
> there, unused.
> 
> >  - more properties lead to more code, more invariants, resulting in
> >    more complex behavior.
> 
> I don't think it's more code. It's less code: instead of writing
> multiple parsers, you have a target-independent parser, and the
> compatibility hacks would be isolated on the legacy property-setters.
> 
> >  - increased maintenance (fixing/rewriting/testing/name it)
> 
> It looks like as much maintenance as having a target-specific
> parse_featurestr() function. Maybe even less, because now we can reuse
> the same parser for the other targets.
> 
> >  - maintenance  will multiply by target amount that has legacy features that
> >    do not map 1:1 into properties
> 
> If a legacy feature do not map 1:1 into a property, you just add a
> legacy property-setter that will translate it into the "true"
> properties. It would contain exactly the same code you would put inside
> an "else if (!strcmp(...)) { ... }" block inside parse_featurestr().
> 
> > 
> > Legacy format converter allows to mitigate or reduce above mention problems:
> >  + keeps CPU API simpler, new interfaces /QMP .../ see only CPU API and
> >    expected not to behave in legacy way.
> 
> The legacy property doesn't even need to be a "static" one, and doesn't
> even need to have a getter. Other code would just never use it. It would
> be similar to the legacy properties registered by
> qdev_property_add_legacy() (maybe we could simply reuse that existing
> mechanism?).
> 
> 
> >  + targets provide simple converter from command line format (could be
> >    painlessly deprecated in future)
> 
> This would have exactly the same functionality: targets provide simple
> list of legacy property-setters to translate old options. But now with a
> target-independent parser.
> 
> 
> >  + allows to centralize point of conversion in one place, using utilities
> >    that target provides if it cares.
> 
> This is the main difference, I guess. But I think not having to
> duplicate parsing code and just have to provide a bunch of
> property-setter functions would be better than having target-specific
> parsing code just to "keep things in one place". You can easily put all
> the legacy setter functions in the same place inside the code, if you
> want to group them into a visible place.
> 
> By the way, I don't think they _must_ be properties. All I want to avoid
> is having to duplicate the full featurestring parser for each target,
> when the only difference is that the targets will handle a few specific
> options differently from -device and QMP. Then featurestr parser could
> be a generic parser that simply set properties. But I would be happy
> with an approch based with something like a:
>   parse_single_feature(const char *key, const char *value)
> callback in CPUClass, called by the generic parser.
> 
> 
> >  + 'using global properties' and 'cpu_name=>class_name conversion in
> >     converter' opens easy road to removing CPUxxxState.cpu_model_str and as
> >     result simplifying CPUxxxState, copy_cpu() and cpu_xxx_init()
> >  + cpu_init() would be easier to unify/simplify and could be reduced to
> >    object_new() + set instance specific properties() + realize()
> >  - practically every QOMified CPU that would use unified cpu_init, should
> >    provide as minimum converter cpu_name=>class_name, but it could be fixed
> >    gradually per target.
> 
> 
> I don't think this has anything to do with the legacy-parse_featurestr
> handling. You can simplify/generalize cpu_init() in exactly the same
> way, the difference is that (wherever you call it) you can call a
> generic featurestr parser, instead of a target-specific one.
> 
> IIRC, other target maintainers expressed interest in allowing options to
> be set in "-cpu". And we already have a feature string parser in x86, we
> just need to make it target-independent so other targets could use it.
> 
> 
> >  - there will be ugly converter callback for a while, but it's temporary,
> >    after it's deprecation users would be expected to use
> >    foo-cpu-class,foo=xxx format
> 
> The same can be said by the legacy property-setter (or a
> parse_single_feature() callback). They could be temporary and would be
> used only if a "-cpu" option that required compatibility translation was
> used.
> 
> 
> > 
> > 
> > In summary I think it would be better to keep CPU object as simple as 
> > possible
> > with a stable API(a property set) for the every interface and convert legacy
> > command-line format into this API.
> > It will keep legacy code contained inside one function, which eventually
> > should disappear as legacy features are deprecated.
> 
> All above said, I am not strongly against your approach, but I believe
> we could try to make the feature string parsing code generic and
> reusable (with target-specific properties or callbacks), instead of
> having to write a new generic feature string parsing function.
> 
> Anyway, the above is just bikeshedding to me, and your approach is an
> improvement, already, even if we try to make the parser more generic
> later.
>
feature string parser in this series isn't meant to be generic, it's just
a simplification of the current function and splitting parsing
features & setting properties into separate steps. kind of what you did with
cpu_name+feature_string split.

It's not clear to me what you mean under legacy properties, could you
throw in a hack using as example xlevel feature to demonstrate what you mean,
please?

> But the questions below about global properties seem important/interesting:
> 
> > 
> > > 
> > > 
> > > > 
> > > >         // with classes and global properties we could get rid of the
> > > > field // cpu_model_str in CPUxxxState
> > > >         return prop_list, cpu_class_name        
> > > >     }
> > > > 
> > > >     foreach (prop_list)
> > > >         add_global_prop(cpu_class_name,prop,val)
> > > > 
> > > >     // could be later transformed to property of board object and
> > > >     // we could get rid of one more global var
> > > >     cpu_model = cpu_class_name
> > > 
> > > The cpu_model string will be immediately used to create the CPU object
> > > just after this code runs. So do we really need to do the parsing before
> > > creating the CPU object and use global properties for that? It looks
> > > like unnecessary added complexity.
> > It might be used immediately or might be not in case of hot-plug or
> > copy_cpu() or depending where converter is called. Ideally there won't be
> > cpu_model string at all.
> 
> I don't see the difference between storing a cpu_model string and
> storing a dictionary to be used later. It would just be a
> micro-optimization for the code that creates new CPUs.
Dictionary is an implementation detail, a temporary list of properties that
would be applied to global properties and then list destroyed. It is there now
because:
 1. it's just a convenient data structure to implement current weird +-foo
    semantic
 2. a necessary intermediate step to hold list of properties because
    properties/features is not yet converted into static properties. Properties
    could/should be pushed directly into global properties if possible (i.e.
    when property is static), then there won't be any need in intermediate
    property list at all.

There is not compelling reason to use dictionary in generic code, it could be
removed or be hidden inside current featurestr parser.

> 
> 
> > 
> > Presently all created CPUs use the same cpu_model string so parsing it
> > only once instead of N times would be already an improvement.
> 
> I don't believe we would want to make the code more complex just to
> optimize string parsing that is run a few times during QEMU
> initialization. That would be unnecessary optimization.
If we move generalized parsing out into vl.c & *-user/main.c option parsing
code, it would simplify every target. I would be better to do option parsing
at the time when options are parsed in a single place and let boards/targets to
concentrate on creating cpus (without caring about options at all, not really
their job).

> 
> > 
> > Pushing common features for type into global properties also makes sense,
> > isn't global properties made specifically for this?
> 
> Yes, but do we want to make "-cpu" affect every single
> -device/device_add call for CPUs, or only the ones created on
> initialization? That's the main question, I believe.
If user overrides default cpu model features with -cpu option it would be only
logical for hot-plugged cpus to use that overrides as well, instead of creating
cpu that would be different (it may confuse guests, to the point of crashing).
IMHO, It would be surprising for user to get a different cpu on hot-plug than
already existing ones, in most cases.

And if user likes to shoot in its own foot, it will be possible to override
global features by adding extra foo=xxx to a specific device_add/-device
command.

> 
> > 
> > If common features are pushed in global properties, cpu hot-plug could look
> > like:
> >    device_add driver=cpu_x86_foo_class,[apicid|id]=xxx
> > and all common features that user has overridden on command line would
> > applied to a new cpu without extra work from user.
> 
> CPU hotplug is an interesting case: if thinking only about the CPUs
> created from the command-line the "-global vs -device" question wouldn't
> make such a big difference. But in the case of device_add for CPU
> hotplug, it would be very different. But I am not sure which option is
> less surprising for users.
> 
> What others think? Andreas? Should "-cpu" affect only the CPUs created
> on initialization, or all CPUs created by -device/device_add for the
> lifetime of the QEMU process?  (in other words, should the options to
> "-cpu" be translated to "-device" arguments, or to "-global" arguments?)
> 
> (BTW about the device_add example above: I believe we will want the
> hotplug interface to be based on a "cpu_index" parameter/property, to
> abstract out the complexity of the APIC ID calculation)
there was other opinion about cpu_index
http://permalink.gmane.org/gmane.comp.emulators.qemu/151626
 
However if board will allocate sockets (Anthony suggested to try links for
this) then user won't have to calculate ID, instead of user will have to
enumerate existing cpu links and use IDs embedded in it for managing
cpus.

> 
> 
> 
> > 
> > > 
> > > > }
> > > > 
> > > > > 
> > > > > (This will probably require calling cpudef_2_x86_cpu() before
> > > > > cpu_x86_parse_featurestr(), and changing the existing
> > > > > cpu_x86_parse_featurestr() code to change the cpu object, not a
> > > > > x86_def_t struct, but I think this will simplify the logic a lot)
> > > > You cannot set/uset +-feat directly on CPU without breaking current
> > > > behavior where -feat overrides all +feat no matter in what order they 
> > > > are
> > > > specified. That's why dictionary is used to easily maintain "if(not 
> > > > feat)
> > > > ignore" logic and avoid duplication. Pls, look at commit
> > > > https://github.com/imammedo/qemu/commit/ea0573ded2f637844f02142437f4a21ed74ec7f3
> > > > that converts +-feat into canonical feat=on/off form.
> > > 
> > > Well, you can make feature parsing code could simply save the
> > > +feat/-feat results internally, and set/unset the properties directly at
> > > the CPU in the end. You can even use a dictionary internally, the point
> > > is that you don't need to expose the dictionary-based interface to the
> > > outside if not necessary (making the API much simpler, IMO).
> > dictionary outside of cpu_x86_parse_featurestr() is temporary,
> > until all features is converted into static properties, after that it could
> > be moved inside cpu_x86_parse_featurestr(), which would push custom 
> > properties
> > into global properties list and be called only once.
> > Additionally it would be possible to remove introduced here
> > cpu_x86_set_props(), simplifying cpu_x86_(init|create)().
> > I'd like to keep legacy handling in one compact place and as much as 
> > possible
> > separated from CPU object itself.
> > 
> 
> I see. If you are going to use global properties, the dictionary
> approach makes a lot of sense.
> 
> 
> > > 
> > > > 
> > > > And if features are applied directly to CPU, it would require to another
> > > > rewrite if global properties are to be used. Which IMHO should be
> > > > eventually done since -cpu ... looks like global parameters for a
> > > > specific cpu type.
> > > 
> > > If we really are going to use global properties for the featurestring
> > > result, we will need to parse the featurestring before creating the CPU
> > > object, then you are correct.
> > > 
> > > I'm questioning the need to use global properties for the
> > > featurestring parsing, if we can simply set the properties after
> > > creating the CPU object. I expect "-cpu" to be easily translatable to
> > > "-device" options, not to "-global" options.
> > > 
> > > In other words, I expect this:
> > > 
> > >   -cpu foo,+xxx,-yyy
> > > 
> > > to be translated in the future into something like:
> > > 
> > >   -device foo-x86-cpu,xxx=on,yyy=off
> > > 
> > > and not to:
> > > 
> > >   -global foo-x86-cpu.xxx=on -global foo-x86-cpu.yyy=off -device 
> > > foo-x86-cpu
> > > 
> > > In other words, I disagree about "-cpu ..." looking like global
> > > parameters for a specific CPU type.
> > I can't agree with you here, on the contrary second option makes more sense
> > for me.
> > But both ways are lead to the same result and both ways will be possible to
> > use.
> 
> Both ways have different results if we think about using device_add for
> CPU hotplug, though.
> 
> > 
> > However if you look at current -cpu cpu_name,foo=xxx,... it is template for
> > all CPUs of type cpu_name, qemu is creating. That's maps perfectly into
> > global properties for a specific type.
> 
> I see it as a template of the CPUs QEMU is creating _at initialization_,
> but I understand that seeing it as a template for all CPUs QEMU will
> ever create makes sense, too. We just have to decide which way to go.
That's what I was trying to explain, I'm sorry my poor English that it took so
many words to get here.


> 
> 
> > 
> > If you are concerned with command line length the for single/several cpus
> > case then syntax "-device foo-cpu,xxx=on,yyy=off -device foo-cpu,.."
> > will be shorter, but with many cpus syntax:
> >  -global foo-cpu.yyy=off -device foo-cpu ... 
> > will be more compact.
> 
> command-line length won't be an issue if users use a config file.  :-)
> 
> > 
> > And as I explained before, cpu hot-plug and copy_cpu() will benefit from 
> > usage
> > of global properties as well.
> > 
> > I think "-cpu foo-cpu,foo=x -smp n  -numa node,cpus=..." could be
> > translated into something like:
> >    -global foo-cpu.foo=x
> >    -device foo-cpu,id=a,node=nodeA
> >    -device foo-cpu,id=b,node=nodeB
> >    ...
> >    or
> >    -device foo-cpu,id=f(cpuN,nodeN)
> >    ...
> > where common options are pushed into global properties and only instance
> > specific ones are specified per instance.
> > 
> > Do you see any issues ahead that global properties would introduce?
> 
> I see no issue, except that it was not what I was expecting at first. I
> just don't know which one is less surprising for users (including
> libvirt).
users shouldn't be affected, -cpu will continue to works the same way until we
start deprecating legacy behavior.

> 
> Summary: I can see that translating "-cpu" to global properties may make
> sense, too. Now I'm divided. I hope to hear what Andreas (and other
> maintainers) think about this.
> 
> [...]
> 
> -- 
> Eduardo
> 


-- 
Regards,
  Igor