Re: [Qemu-devel] [PATCH v11 0/8] Add a generic loader

[Markus Armbruster]

Paolo Bonzini <address@hidden> writes:

> On 22/09/2016 11:19, Markus Armbruster wrote:
>>> > I think -device is okay for something that isn't a "backend" but is
>>> > directly guest-visible.
>> Well, the contents of a block device is just as guest-visible.  We split
>> the device in a frontend and a backend, and the contents comes from the
>> backend.
>> 
>> We traditionally don't model memory as a split device.  Perhaps we
>> should.  Regardless of whether we actually do, "contents of a memory
>> device that you need to create by some other means (explicit or
>> implicit)" feels much more like -object than like -device to me.
>
> The closest precedents here are "-bios" (not an object at all), PCI ROMs
> (a property points to the file), and "-pflash" (split into backend and
> frontend, the frontend being a device).

-pflash ARG is sugar for -drive if=pflash,file=ARG.  Like all -drive
other than drive if=none, it creates a block backend and asks the board
to create a block frontend.  Some boards create a modern QOMified device
such as "cfi.pflash01", some create just a ROM memory region, which
isn't a device (in the same way RAM normally isn't), and some ignore the
request silently.

The modern, fully general form for QOMified devices would be to define
the backend with -drive if=none, and the frontend with -device.

-bios ARG is sugar for -machine firmware=ARG.  This asks the board to
load firmware from file ARG.  Again, boards could create a modern
QOMified ROM device (theoretical, not sure there are any), a non-device
ROM memory region, or ignore the request.  The difference to pflash is
that we don't wrap a full-blown block backend around the file, simply
because that would complicate things for no gain.  But that's detail;
it's playing the role of a backend all the same.

A (QOMified) PCI may have a PCI ROM.  To configure its contents, you use
a PCI device property naming a file.

In all cases, QOMified devices get bits from the host via a backend.
The backend can be an explicit object (e.g. a block backend) or, if the
backend is trivial, folded into the frontend.

> I think there is a device concept in here, the question is whether you
> want to split the backend and frontend.  For read-only data the
> precedents favor not splitting it.

Yes, we sometimes choose to fold trivial backends into the frontends
instead of doing a split device.  That's fine.

However, in the case we're discussing, we're not doing that!  There is
no RAM device with a trivial backend folded in.  There's only a weird
pseudo-device that copies the contents of a file into memory, then sits
around doing absolutely nothing.

For me, that's an abuse of -device.  It's certainly no worse than some
other parts of the QEMU command line / QMP.  Not exactly a good excuse,
though :)