Re: [PATCH v3 03/12] vfio-user: define vfio-user-server object

[Markus Armbruster <armbru@redhat.com>]

Kevin Wolf <kwolf@redhat.com> writes:

> Am 05.11.2021 um 11:08 hat Markus Armbruster geschrieben:
>> Kevin Wolf <kwolf@redhat.com> writes:
>> 
>> > Am 04.11.2021 um 13:13 hat Markus Armbruster geschrieben:
>> >> The old syntax almost always has its quirks.  Ideally, we'd somehow get
>> >> from quirky old to boring new in an orderly manner.  Sadly, we still
>> >> don't have good solutions for that.  To make progress, we commonly
>> >> combine JSON new with quirky old.
>> >> 
>> >> qemu-system-FOO -object works that way.  object_option_parse() parses
>> >> either JSON or QemuOpts.  It wraps the former in a QObject visitor, and
>> >> the latter in an opts visitor.
>> >> 
>> >> QemuOpts is flat by design[*], so the opts visitor parses flat QemuOpts
>> >> from a (possibly non-flat) QAPI type.  How exactly it flattens, and how
>> >> it handles clashes I don't remember.
>> >> 
>> >> Sadly, this means that we get quirky old even for new object types.
>> >
>> > For -object in the system emulator (the tools all use the keyval
>> > visitor, so there it would work as expected), the only reason that we
>> > need to keep the quirky old code path around is the list handling in
>> > memory-backend.host-nodes.
>> >
>> > The main difficulty there is that the old QemuOpts based code path
>> > allows specifying the option twice and both of them would effectively be
>> > combined. Do we have any idea how to replicate this in a keyval parser
>> > based world?
>> 
>> I can see just two clean solutions, but both involve upending a lot of
>> code.
>> 
>> We can fuse keyval parser and visitor to get a schema-directed parser.
>> 
>> We can change the abstract keyval syntax to permit repeated keys.  This
>> means replacing QDict in in the abstract syntax tree, with fallout in
>> the visitor.
>> 
>> Even if we find a practical solution, I don't like the combination of
>> "you may give the same parameter multiple times, and the last one wins"
>> and "for a list-valued parameter, the values of repeated parameters are
>> collected into a list".  Each makes sense on its own.  The combination
>> not so much.  Inheriting "last one wins" from QemuOpts may have been a
>> mistake.
>> 
>> The keyval way of doing lists (inherited from the block layer's usage of
>> dotted keys?  I don't remember) requires the user to count, which isn't
>> exactly nice, either.
>
> Yes. If we didn't have to maintain compatibility (or actually as soon as
> we degrade non-JSON option lists to HMP-level support), I would
> introduce [] and {} syntax for lists and dicts, even if that means that
> use of these characters in strings doesn't work any more or only in a
> limited way. I think this would be the best compromise for usability.
>
> Anyway, this doesn't help us with the compatibility problem we're
> discussing here.
>
>> > If not, do we want to use the remaining time until 6.2 to deprecate
>> > this? The nasty part is that the only syntax that works both now and in
>> > the future is JSON. We can't easily accept the new keyval syntax while
>> > still using the QemuOpts based code.
>> 
>> What exactly do you propose to deprecate?
>
> We can deprecate on two different levels. I think it's useful to do
> both:
>
> 1. Broad deprecation: Stable non-JSON interfaces are degraded to
>    a HMP-like compatibility promise.

Calling it "deprecation" might be confusing.  HMP isn't deprecated, it's
merely not a stable interface.  That's kind of like "deprecated when you
need stable", but saying "not a stable interface" is clearer.

When I write "deprecate" below, I mean something like "go use something
else (no conditions)".  When I mean "use something else when you need
stable", I write "degrade" (short for "degrade to an HMP-like
compatibility promise").

>                                      Obviously, this can only be done
>    for options that support JSON.

We can also degrade or even deprecate sugar options in favor of the real
ones.  Case by case, I guess.

>                                   Peter Maydell also wants to do this
>    only after a big user (read: libvirt) has implemented and is
>    using JSON, basically as a proof that the alternative is working.
>
>    So this can certainly be done for -object. I believe libvirt also
>    uses JSON for -device now, so this should be fine now, too.

The non-sugar options supporting JSON are -audiodev, -blockdev, -compat,
-display (partially), -machine (I think), -object.

-netdev is QAPIfied, but still uses QemuOpts.  Too late for 6.2, I'm
afraid.

>                                                                Possibly
>    -drive (in favour of -blockdev), though I'm not completely sure if we
>    have gotten rid of the final users of -drive. (CCing Peter Krempa for
>    details.)

The problem with deprecating -drive is configuring onboard block
devices.  We need a stable interface for that, and it must be usable
together with -blockdev.

We provided such an interface (machine properties) for some onboard
block devices starting with commit ebc29e1bea "pc: Support firmware
configuration with -blockdev".  Many more remain, I believe.

>    This degradation of the compatibility promise doesn't tell users what
>    exactly is going to change, which is why doing the second one, too,
>    might be nice.
>
> 2. Narrow deprecation: We can just deprecate the non-JSON form, or
>    certain aspects of it, of memory-backend.host-nodes. This is the
>    specific things that stops us from switching -object to keyval.
>
>    a. Deprecate the whole option. If you want to use it and need a
>       stable interface, you have to use JSON. We'll just switch the
>       non-JSON form on a flag day. Before it, you need to use QemuOpts +
>       OptsVisitor syntax for the list; after it, you need to use keyval
>       syntax.

I parse "the whole option" as "-object with dotted keys argument".
Correct?

>    b. Deprecate only repeating the option. memory-backend is changed to
>       first try visiting a list, and if that fails, it visits a string
>       and goes through a string visitor locally to keep supporting the
>       integer range syntax.

Possible problem: integer range syntax must not leak into the JSON form.

>    c. Deprecate all list values, but keep supporting a single integer
>       value by using an alternate between list and int.

Single int should also not leak into JSON.

> Picking one of these four options is enough to convert -object to
> keyval. I would suggest doing both 1. and one of the options in 2.

I'm grateful for your analysis.