Re: [Qemu-devel] [RFC PATCH 1/5] qdev: Create qdev_get_dev_path()

[Markus Armbruster]

Jan Kiszka <address@hidden> writes:

> Markus Armbruster wrote:
>> Jan Kiszka <address@hidden> writes:
>> 
>>> Markus Armbruster wrote:
>>>> Paul Brook <address@hidden> writes:
>>>>
>>>>> Alex Williamson <address@hidden> writes:
>>>>>
>>>>>> On Mon, 2010-06-14 at 08:39 +0200, Markus Armbruster wrote:
>>>>>>> Could you explain why you add "identified properties of the immediate
>>>>>>> parent bus and device"?  They make the result ver much *not* a "dev
>>>>>>> path" in the qdev sense...
>>>>>> In order to try to get a unique string.  Without looking into device
>>>>>> properties, two e1000s would both be:
>>>>>>
>>>>>> /main-system-bus/pci.0/e1000
>>>>>> /main-system-bus/pci.0/e1000
>>>>>>
>>>>>> Which is no better than simply "e1000" and would require us to fall back
>>>>>> to instance ids again.  The goal here is that anything that makes use of
>>>>>> passing a dev when registering a vmstate gets an instance id of zero.
>>>>> You already got the information you need, you just put it in the wrong 
>>>>> place. 
>>>>> The canonical ID for the device could be its bus address. In practice 
>>>>> we'd 
>>>>> probably want to allow the user to specify it by name, provided these are 
>>>>> unique. e.g. in the above machine we could accept [...]/virtiio-blk-pci 
>>>>> would 
>>>>> as an aias for [...]:_09.0. Device names have a restricted namespace, so 
>>>>> we 
>>>>> can use an initial prefix to disambiguate a name/label from a bus address.
>>>>>
>>>>> For busses that don't have a consistent addressing scheme then some sort 
>>>>> of 
>>>>> instance ID is unavoidable. I guess it may be possible to invent 
>>>>> something 
>>>>> based on other device properties (e.g. address of the first IO 
>>>>> port/memory 
>>>>> region).
>>>> When that's inconvenient or impossible, we can still punt to user: make
>>>> device ID mandatory.
>>> No option due to auto-created devices. And auto-generating IDs would
>>> just create usability issues.
>> 
>> Auto-generated IDs would become part of the ABI.  Really so bad that
>> it's "no option"?  Mind, device ID becomes mandatory *only* for devices
>> that don't have a useful bus address.  We could even waive the ID
>> requirement for the first device of a kind, i.e. require ID if and only
>> if it's needed to disambiguate.
>
> IDs are there to find devices the user (or a higher level tool) passed
> to QEMU, qtree paths allow to locate _every_ device in a VM, and that in
> a well-organized hierarchy. That allows to explore and address a qtree
> element at the same time.
>
>> 
>>>> We obviously need a way to unambigously name a device.  It's okay to
>>>> have multiple names for the same device.
>>>>
>>>> If the device has a device ID, that's an unambigous name.
>>>>
>>>> qdev paths may be ambigous when path components are resolved to driver
>>>> names instead of IDs.
>>>>
>>>> Alex proposed to disambiguate by adding "identified properties of the
>>>> immediate parent bus and device" to the path component.  For PCI, these
>>>> are dev.fn.  Likewise for any other bus where devices have unambigous
>>>> bus address.  The driver name carries no information!
>>> >From user POV, driver names are very handly to address a device
>>> intuitively - except for the case you have tones of devices on the same
>>> bus that are handled by the same driver. For that case we need to
>>> augment the device name with a useful per-bus ID, derived from the bus
>>> address where available, otherwise based on instance numbers.
>> 
>> I'm not arguing against the use of driver names at all.
>> 
>>>> For other buses, we need to make something up.
>>>>
>>>> Note that addressing by bus address rather than name is generally
>>>> useful, not just in the context of savevm.  For instance, I'd appreciate
>>>> being able to say something like "device_del pci.0/04.0".
>>> And I prefer "device_del [.../]pci.0/e1000". Otherwise you need to dump
>>> the bus first before you can identify which device you want to remove.
>> 
>> It's not either/or.  Addressing by ID continues to work.  Addressing by
>> bus/driver-name continues to work.  We merely add addressing by
>> bus/@bus-address.
>
> The format I will propose is "global-ID|/absolute/path", no more
> /path/global-ID as this comes with the risk of ambiguity (ID may shadow
> bus-local name of a device).

Doesn't this break existing usage?

We have a rule to resolve any ambiguity added by ID: it always takes
precedence over driver name.  What path/ID does add is shadowing: it can
make a device inaccessible by driver name.  Not much of a difference to
adding a second device with the same driver name.

>>>> An easy way to get that is to reserve part of the name space for bus
>>>> addresses.  If the path component starts with a letter, it's an ID or
>>>> driver name.  If it starts with say '@', it's a bus address in
>>>> bus-specific syntax.  The bus provides a method to look it up.
>>> I would prefer <driver>[@<bus-address>|.<instance-no>]. The former is
>>> set for buses that implement some to-be-defined device addressing
>>> service, the latter is the default on buses where that service is not
>>> available.
>> 
>> I object to <driver>@<bus-address>, because the <driver> part carries no
>> information.
>
> I does for a human being as bus addresses tend to be unreadable and can
> easily be confused, hence the additional, sometimes redundant driver name.

I *strenuously* object to making the driver name mandatory with
bus-address addressing.

If you think the human user needs to be protected from mistakes by
making him supply redundant information, then "device_del pci.0/04.0
e1000" is much better way than complicating device paths for human and
machine users alike, not to mention uses internal to QEMU.

>> Not the case for <driver>.<instance-no>.  We still need a suitable
>> definition of <instance-no>.  Possible definitions:
>> 
>> * n-th creation of a <driver> device.  Drawbacks: depends on creation
>>   order.  Relatively hard to maintain across migration.
>> 
>> * n-th instance of a <driver> device.  Drawback: changes on unplug.
>>   Good enough for interactive use, but it doesn't provide a stable
>>   device name.
>
> Every hotplug-capable bus must have a proper addressing scheme, I think
> this is a reasonable and achievable requirement. Then we don't need
> instance numbers for those buses.

What about USB?

>> When counting <driver> devices either way, we can count per bus or
>> globally.  I prefer per bus.
>
> Yes, counting should be both per-driver and per-bus ("the <n>th device
> managed by <driver> on this bus").

Works for me.

>> None of the above instance numbers are nearly as neat as bus addresses.
>
> Right, wherever they are available.
>
> Jan