Re: [Qemu-devel] [PATCH v3 0/7] Improvements and clean-ups related to -net

[Thomas Huth <thuth@redhat.com>]

On 21.02.2018 11:41, Mark Cave-Ayland wrote:
> On 21/02/18 10:18, Thomas Huth wrote:
> 
>> "-net" is a legacy option that often causes confusion and
>> misconfigurations for the users, since most people are not aware
>> of the underlying "vlan" (i.e. hub) concept that is used for this
>> parameter. The prefered way of configuring your network stack is
>> to use "--netdev" instead, which gives you a clean 1:1 connection
>> between your emulated guest hardware and the host network backend.
>>
>> However, there are two reasons why we could not completely deprecate
>> "-net" yet:
>>
>> 1) Convenience:
>> In some cases, it's more convenient to use "-net" instead of "--netdev",
>> e.g. if you just want to have a "tap" network connection, it's faster
>> to type "-net nic -net tap" instead of "--device e1000,netdev=n1
>> --netdev tap,id=n1".
>>
>> 2) On-board NICs:
>> Currently the "-net nic" parameter is the only way to configure on-
>> board NICs on certain (embedded) machines via the nd_table[] array.
>>
>> So beside some generic clean-ups and removal of code that has been
>> marked as deprecated since QEMU 2.10 already, this patch series intro-
>> duces a new parameter "--nic" (in patch 7) which should be able to re-
>> place "-net" in the long run completely: This new convenience parameter
>> can be used to configure the default/on-board guest HW together with a
>> host network backend in a very compact way. To configure a tap backend
>> for the default NIC, you just have to type "--nic tap" here for example.
> 
> Hi Thomas,
> 
> This is a great improvement for configuring on-board NICs! I do have a
> couple of questions about your patchset based on personal experience:
> 
> 1) Does the new -nic syntax support multiple on-board NICs? I remember
> seeing this on some of the ARM boards I was studying when trying to
> implement something similar for SPARC.

Yes, that's possible. If you specify the "--nic" parameter multiple
times, it populates multiple entries in the nd_table[] array, so boards
can pick up multiple NICs from there. For example hw/arm/xlnx-zynqmp.c
initializes up to 4 (XLNX_ZYNQMP_NUM_GEMS) on-board NICs this way.

> 2) Is it possible to provide a convenient wrapper function to handle the
> logic related to determining whether a specified NIC is on-board or not?
> 
> For example take a look at
> https://git.qemu.org/?p=qemu.git;a=blob;f=hw/sparc64/sun4u.c;h=da28ab9413efdbe0bf0e1d3bf2b545577b83d88a;hb=a6e0344fa0e09413324835ae122c4cadd7890231#l596
> where we have a nice dance to check whether the device specified is
> on-board or not, and if it isn't to plug it in as an extra PCI device
> instead. The complication here is that nd_tables[] is populated by
> default (and also with -net nic -net user) but not when someone
> specifies a single card via -netdev.
> 
> From memory this was derived from the code in
> https://git.qemu.org/?p=qemu.git;a=blob;f=hw/arm/realview.c;h=87cd1e583cd20b6d8a2beeef1cba6977496d4477;hb=a6e0344fa0e09413324835ae122c4cadd7890231#l259
> if having another example helps?

Do you have something in mind how the interface to such a function
should look like?
Not sure whether there is a easy, magic solution for all boards ...
there are multiple problems, e.g. that you can normally not specify
on-board NICs via --device yet...
Anyway, that would be subject to another clean-up patch series - this
series here only focuses on the net/* files, not on the files in the hw/
folder.

 Thomas