Re: [Qemu-devel] [PATCH v3 1/6] hw/misc/dyn_sysbus_binding: helpers for sysbus device dynamic binding

[Paolo Bonzini <pbonzini@redhat.com>]

Il 10/09/2014 15:51, Eric Auger ha scritto:
> On 09/10/2014 12:09 PM, Alexander Graf wrote:
>>
>>
>> On 10.09.14 12:05, Paolo Bonzini wrote:
>>> Il 10/09/2014 11:56, Alexander Graf ha scritto:
>>>>
>>>>
>>>> On 10.09.14 11:43, Paolo Bonzini wrote:
>>>>> Il 10/09/2014 11:31, Alexander Graf ha scritto:
>>>>>>>> Yeah, please do the registration in sysbus.c, not in virt.c.  There is
>>>>>>>> no reason to make the platform_bus_init_notify+DynSysbusNotifier
>>>>>>>> interface public.  The code in sysbus.c can fill in the fields.
>>>>>> Sysbus != Platform bus. Sysbus is an in-QEMU representation of a
>>>>>> pseudo-bus that we put all devices onto that we consider unsorted.
>>>>>>
>>>>>> Platform bus is a machine representation of an actual bus that devices
>>>>>> are attached to. These devices usually are sysbus devices.
>>>>>
>>>>> Is there any difference between the two?
>>>>>
>>>>> Take a machine that has two chips, a SoC that does everything except
>>>>> USB, and a USB controller chip.
>>>>>
>>>>> Strictly speaking the USB controller chip would be on a "platform bus",
>>>>> but we would likely put it on sysbus.
>>>>>
>>>>> Why should it matter whether the devices are static or dynamic, for the
>>>>> sake of calling something the "system" or the "platform" bus?  I would
>>>>> say that QEMU calls "sysbus" the platform bus.
>>>>>
>>>>> Some devices (e.g. the local APIC in x86, or the in-core timers and GIC
>>>>> in ARM) should probably not be in sysbus at all, and should attach
>>>>> directly to the CPU address space.  But that is a quirk in the modeling
>>>>> of those devices, it shouldn't mean that sysbus is not a "platform" bus.
>>>>
>>>> On e500 for example, we have a predefined CCSR region. That is a machine
>>>> defined "platform bus". The offsets inside that region are strictly
>>>> defined by the spec.
>>>>
>>>> Now take the serial ports. We have space for 2 serial ports inside of
>>>> that CCSR region. We can spawn these 2 ports in the machine file based
>>>> on -serial, but if you want to spawn them with -device, how do you tell
>>>> the machine whether they should go into the "big bucket platform bus" or
>>>> the "CCSR platform bus"?
>>>
>>> Two possibilities:
>>>
>>> 1) you would use two instances of sysbus (one default, one created by
>>> the board) and specify ",bus=ccsr" on the command line when you want to
>>> add the device to the CCSR region.
>>>
>>> The two would work exactly the same way, only with different algorithms
>>> for resource allocation.
>>>
>>> 2) similar to ISA, you would create a new ccsr-bus device and a new
>>> ccsr-serial device, and use -device ccsr-serial,index=[0|1],chardev=foo
>>> to specify which of the two serial ports this is for.  Most of the fdt
>>> magic could be shared by the sysbus and CCSR cases.
>>>
>>> I think I prefer (2)...
>>
>> Fair enough.
>>
>> As far as moving "platform bus" logic into sysbus, I'd really like to
>> hold back and see what this whole thing ends up getting used for first.
>>
>> So for now, I'd definitely prefer to keep "platform bus" logic and
>> "sysbus" logic separate. If we realize that every user only ever uses
>> the dynamic sysbus creation in conjunction with our "platform bus"
>> implementation, we can merge them.
> 
> Hi Paolo, Alex,
> 
> I understand I keep the code in a separate module from sysbus.c. Is that
> the shared conclusion?

I don't think so, but maybe I misunderstood what Alex wrote.

Paolo