Re: [PATCH net-next-2.6] tc: report informations for multiqueue devices

[Patrick McHardy <kaber@trash.net>]

Eric Dumazet wrote:
> Patrick McHardy a écrit :
>> Eric Dumazet wrote:
>>> [PATCH net-next-2.6] tc: report informations for multiqueue devices
>>>
>>> qdisc and classes are not yet displayed by "tc -s -d {qdisc|class} show"
>>> for multiqueue devices.
>>>
>>> We use a new TCA_QINDEX attribute, to report queue index to user space.
>>> iproute2 tc should be changed to eventually display this queue index as in :
>>>
>>> $ tc -s -d qdisc
>>> qdisc pfifo_fast 0: dev eth0 root bands 3 priomap  1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
>>>  Sent 52498 bytes 465 pkt (dropped 0, overlimits 0 requeues 0)
>>>  rate 0bit 0pps backlog 0b 0p requeues 0
>>> qdisc pfifo_fast 0: dev eth0 qindex 1 root bands 3 priomap  1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
>>>  Sent 0 bytes 0 pkt (dropped 0, overlimits 0 requeues 0)
>>>  rate 0bit 0pps backlog 0b 0p requeues 0
>> This might confuse existing userspace since the handle is not unique
>> anymore. libnl f.i. will treat all but the first root qdisc as an
>> update and use it to update the state of the first one. There's also
>> no combined view for applications unaware of multiqueue.
>>
>> Please have a look at the mail I just wrote for some possible ways
>> around this.
> 
> Hum, how can we combine infos on qdisc/class if in the future we allow each queue index
> to have its own qdisc/classes ?
> 
> htb on queue index 0
> cbq on queue index 1

My suggestion was to only dump the statistics in the combined
view and use a virtual qdisc, something like:

qdisc multiqueue 0: dev eth0 root queues 8
  Sent ...
  rate ...

and show each real qdisc as child of this qdisc:

qdisc pfifo_fast <unique handle> dev eth0 parent 0: bands 3 ...
qdisc pfifo_fast <unique handle> dev eth0 parent 0: bands 3 ...

Configuration would be symetrical to this:

tc qdisc add dev eth0 handle 0: root multiqueue
tc qdisc add dev eth0 handle x: parent 0: pfifo_fast
...

without the virtual multiqueue qdisc, the root qdisc would simply
be shared among all queues as today.

> Combining info would lock us and not allow for special configurations.
> Say 
>    macvlan device 0 mapped to queue index 0
>    macvlan device 1 mapped to queue index 1...

Why not?

> For old apps, just give informations for queue 0 as we do now, and
> allow kernel to give more informations only if new application provided a TCA_INDEX attribute
> in its request ?
> 
> (-1 : all queue indexes,  >=0 for a given queue index)

If we don't combine the information, existing multiqueue unaware
applications will get incorrect information. There's also the
problem of non-unique handles, I think we should encode the queue
index in the handle instead of using a new attribute. This needs
a bit more thought though to avoid clashes with user-defined handles.