Re: [Qemu-devel] [PATCH v2] QMP: Introduce the documentation for query-qdm

[Markus Armbruster]

Luiz Capitulino <address@hidden> writes:

> On Tue, 13 Jul 2010 18:20:20 +0200
> Markus Armbruster <address@hidden> wrote:
>
>> Luiz Capitulino <address@hidden> writes:
>> 
>> > On Tue, 13 Jul 2010 13:49:46 +0200
>> > Markus Armbruster <address@hidden> wrote:
>> >
>> >> [cc: kraxel]
>> >> 
>> >> I didn't get around to review v1.  Sorry.
>> >> 
>> >> Miguel Di Ciurcio Filho <address@hidden> writes:
>> >> 
>> >> > Changelog from v1
>> >> > -----------------
>> >> > - renamed "props" to "properties"
>> >> > - updated the examples
>> >> > - reworded the explanations of "name" and "description"
>> >> > - split "type" into a json-object, adding "qmp" and "qdev"
>> >> > - list all possible values for "bus"
>> >> > - list all possible values for "qdev" on "type"
>> >> > - list all possible values for "qmp" on "type"
>> >> > ---
>> >> >  qemu-monitor.hx |   88 
>> >> > +++++++++++++++++++++++++++++++++++++++++++++++++++++++
>> >> >  1 files changed, 88 insertions(+), 0 deletions(-)
>> >> >
>> >> > diff --git a/qemu-monitor.hx b/qemu-monitor.hx
>> >> > index 2af3de6..0055d0a 100644
>> >> > --- a/qemu-monitor.hx
>> >> > +++ b/qemu-monitor.hx
>> >> > @@ -2490,6 +2490,94 @@ STEXI
>> >> >  show device tree
>> >> >  @item info qdm
>> >> >  show qdev device model list
>> >> > +ETEXI
>> >> > +SQMP
>> >> > +query-qdm
>> >> > +---------
>> >> > +
>> >> > +Describe the capabilities of all devices registered with qdev.
>> >> > +
>> >> > +The returned output is a json-array, each element is a json-object 
>> >> > describing
>> >> > +a single device type.
>> >> > +
>> >> > +Each json-object contains the following:
>> >> > +
>> >> > +- "name": name of the device (json-string)
>> >> > +- "bus": the name of the bus type for the device (json-string)
>> >> > +    - Possible values: PCI, SCSI, I2C, ISA, SSI, USB, 
>> >> > virtio-serial-bus, System
>> >> 
>> >> Missing: IDE (hw/ide/qdev.c) and s390-virtio (hw/s390-virtio-bus.c).
>> >
>> > What about "System", is it ok to expose it?
>> 
>> What makes it special?
>
> Not sure I got what you meant.

What makes bus "System" different from other buses?  Unless there's
significant difference, there's no reason to treat it specially.

>> >> > +- "alias": an alias by which the device is also known (json-string, 
>> >> > optional)
>> >> > +- "description": description of the device  (json-string, optional)
>> >> > +- "creatable": whether this device can be created on command line 
>> >> > (json-boolean)
>> >> 
>> >> "on the command line" is misleading; it applies to monitor (human & QMP)
>> >> as well.
>> >> 
>> >> "by the user"?
>> >
>> > When is a device not "creatable"?
>> 
>> Stuff like i440FX: if your board comes with it, you already got it, if
>> not, you can't have it.
>
> Ok, so it's about devices that can be dynamic created and hot plugged, right?

It's not about hot-plug.  For instance, isa-serial is "creatable", but
can't be hot-plugged.  -device can cold-plug it.  device_add doesn't
know how to cold-plug, but that's fixable.

> In this case I think the text should read something like:
>
>  "the device can be dynamically created"

"dynamically" suggests hot-plug.

>> >> > +- "properties": a list where each element is an json-object that 
>> >> > describes a
>> >> > +  property of the device. Each json-object contains the following:
>> >> > +     - "name": the name of the property (json-string)
>> >> > +     - "type": a json-object that contains the following:
>> >> > +        - "qdev": the internal name of the type of the property 
>> >> > (json-string)
>> >> > +            - Possible values: uint8, uint16, uint32, uint64, int32, 
>> >> > macaddr,
>> >> > +              drive, chr, string, netdev, bit, taddr
>> >
>> > Didn't see this before, but we should not use too short names like "chr" 
>> > and
>> > "taddr".
>> >
>> >> > +        - "qmp": the json equivalent type of the internal type 
>> >> > (json-string)
>> >> > +            - Possible values: integer, string, boolean
>> >> 
>> >> Fairly close to JSON Schema, but there are differences.
>> >> 
>> >> Do we need "qdev"?  Is exposing it wise?  Smells a bit too much of
>> >> internal detail for comfort...
>> >
>> > Agreed, although this was a suggestion from me.
>> >
>> >> Could we use "type" just like JSON Schema?  Drop "qdev" or move it out
>> >> of "type", then make "type" what its member "qmp" is now.
>> >
>> > The most important point here is to try to figure out how clients are
>> > going to use this. Actually, the fact that libvirt doesn't actually
>> > need this makes me wonder if we should have it in the first place, at
>> > the same time listing the property name w/o saying what type it accepts
>> > doesn't seem to be that useful either.
>> 
>> Adding stuff we need when we need it is easier and less painfull than
>> taking out stuff when we realize we shouldn't have added it :)
>> 
>> > Let's look at all possibilities:
>> >
>> >  1. Drop "qmp" key and rename "qdev" to "type" (ie. use the qdev type only)
>> >
>> >     This was the original proposal. I don't like it because its values
>> >     are quite low-level and have no meaning in json.
>> >
>> >     If this is needed, then I think it needs a complement specifying the
>> >     json type for it.
>> >
>> >  2. Drop "qdev" key and rename "qmp" to "type" (ie. use the json type only)
>> 
>> Let's start with this, and add more information as needed.
>
> I would be ok with that, I guess.
>
>> >     This is what makes sense when self-documenting the protocol, however,
>> >     given the original proposal, I wonder if this is going to be enough.
>> >
>> >     That is, say a property takes a MAC address. By using the json type
>> >     we will just say "string", how is the client going to know that it's
>> >     actually a MAC address in a string format?
>> 
>> I think the primary use for self-documentation is to let clients figure
>> out whether something they know how to use is actually there.  
>
> Then documenting types is not needed.

... for this particular use case.  There are others.

> I think it's beyond that, one possible advanced usage is to allow clients
> to automatically catch up with protocol additions.
>
> If we add only the json type, then client writers will have to know beforehand
> the qdev type of all properties, which implies that clients will have to be
> modified to learn about new properties.
>
> If the problem is the key's name ("qdev"), then we could have:
>
> - "type": a json-object with type information, contains the following:
>    - "json": the JSON type of the property, to be used when setting it
>              through QMP (json-string)
>    - "format": only present when the property requires special formatting
>                (json-string, optional)
>                  Possible values: macaddr, drive, chardev, netdev, taddr
>
> Now, I only know the format of a macaddr and am assuming that the others
> need special formatting too.

enum PropertyType has little to do with JSON, and should not be exposed
here.  Its purpose is to declare how to parse the argument value from a
*string* (think -device) and store the resulting value in the
DeviceState.

Aside: device_add actually converts all property values to string.
That's because -device came first, and it takes only strings.  Not nice.

What you seem to have in mind is how to give clients more information
about acceptable argument values than just "type".  I agree that such
information may be useful for some applications.  But PropertyType is
not the solution, not even a partial solution (I can elaborate if you
want).  Check out JSON Schema for an attempt at a more complete
solution.

Let's do just the obvious "type" (json-string) now.

[...]