Re: [Qemu-devel] [PATCH v3 6/9] rocker: add new rocker switch device

[Scott Feldman <sfeldma@gmail.com>]

On Fri, Jan 16, 2015 at 1:15 AM, Jason Wang <jasowang@redhat.com> wrote:
>
> On 01/11/2015 11:57 AM, sfeldma@gmail.com wrote:
>> Each port is a netdev and can be paired with using -netdev id=<port name>.
>>
>> Signed-off-by: Scott Feldman <sfeldma@gmail.com>
>> Signed-off-by: Jiri Pirko <jiri@resnulli.us>
>> ---
>
> Looks like the devices does not support state saving. How about tagging
> it with unmigratable first?

State saving would be nice to add in the future.  How do we tag is
unmigratable for now?

>> +    uint64_t switch_id;          /* switch id */
>
> Looks like the function of name ad switch_id are duplicated?

Don't follow...

>> +    if (iovcnt) {
>> +        /* XXX perform Tx offloads */
>> +        /* XXX   silence compiler for now */
>
> Need add TSO here (e1000 is a good reference) or disable TSO in driver.

TSO is not enabled in driver, currently.  This is lower priority now
since the I/O path for traffic to/from the guest CPU is not
performance critical.  The forwarding path offloaded by the device is
the hot path, but that doesn't involve I/O to guest CPU, so Tx/Rx
offloads aren't in play.

But, someday, it would be nice to enable Rx/Tx offloads.

>> +    rocker_tlv_put_le32(buf, &pos, ROCKER_TLV_EVENT_TYPE,
>> +                        ROCKER_TLV_EVENT_TYPE_LINK_CHANGED);
>> +    nest = rocker_tlv_nest_start(buf, &pos, ROCKER_TLV_EVENT_INFO);
>> +    rocker_tlv_put_le32(buf, &pos, ROCKER_TLV_EVENT_LINK_CHANGED_PPORT, pport);
>> +    rocker_tlv_put_u8(buf, &pos, ROCKER_TLV_EVENT_LINK_CHANGED_LINKUP,
>> +                      link_up ? 1 : 0);
>
> Looks like those types are not documented.

Ok, I'll double check spec to make sure we're covered, for the items you found.

>> +    rocker_tlv_put_le16(buf, &pos, ROCKER_TLV_RX_FLAGS, rx_flags);
>> +    rocker_tlv_put_le16(buf, &pos, ROCKER_TLV_RX_CSUM, rx_csum);
>
> Note: Some backend (e.g tap) can do offloading, may consider to add the
> support in the future.

Yup, on TODO list for Rx/Tx offloads.

> Using only 1 queues is ok for multiqueue backend. In the future may
> consider to use all.

When we get to QoS for working, I think we'll want to enable these
multiqueues.  QoS support is not implemented in rocker, currently, but
there is support defined in the OF-DPA spec as part of ACL processing.
So this is future work.

> Thanks

Thanks for reviewing.