On 2020-06-09 14:58, Jakub Kicinski wrote:
On Mon,  8 Jun 2020 17:00:59 -0700 David Wilder wrote:
Hyper-visors owning a PF are allowed by Emulex specification to provide
a VF with separate physical and/or logical link states. However, on
linux, a VF driver must chose one or the other.

My scenario is a proprietary driver controlling the PF, be2net is the VF.
When I do a physical cable pull test the PF sends a link event
notification to the VF with the "physical" link status but this is
ignored in be2net (see be_async_link_state_process() ).

The PF is reporting the adapter type as:
0xe228   /* Device id for VF in Lancer */

I added a module parameter "use_pf_link_state". When set the VF should
ignore logical link state and use the physical link state.

However I have an issue making this work.  When the cable is pulled I
see two link statuses reported:
[1706100.767718] be2net 8002:01:00.0 eth1: Link is Down
[1706101.189298] be2net 8002:01:00.0 eth1: Link is Up

be_link_status_update() is called twice, the first with the physical link status called from be_async_link_state_process(), and the second with the
logical link status from be_get_link_ksettings().

I am unsure why be_async_link_state_process() is called from
be_get_link_ksettings(), it results in multiple link state changes
(even in the un-patched case). If I eliminate this call then it works.
But I am un-sure of this change.

Signed-off-by: David Wilder <dwil...@us.ibm.com>

Hm. Just looking at the code in __be_cmd_set_logical_link_config():


        if (link_state == IFLA_VF_LINK_STATE_ENABLE ||
            link_state == IFLA_VF_LINK_STATE_AUTO)
                link_config |= PLINK_ENABLE;

        if (link_state == IFLA_VF_LINK_STATE_AUTO)
                link_config |= PLINK_TRACK;

Maybe we shouldn't set ENABLE for AUTO?

If I am understanding this correctly, this is used by the linux PF driver to configure
how link status is delivered to a VF.

My problem is one of interoperability between the PF (not linux) and the VF is running on linux. The PF driver is implemented to the Emulex/Broadcom spec, which allows a PF driver to be configured such that the VF can be notified of both physical and logical link status, separately.


The module parameter is definitely not a good idea, what you're asking
for seems to be well within the expectation from the
.ndo_set_vf_link_state config, so it seems the driver / firmware is just
not implementing that right.

I am attempting to resolve an issue that the linux implementation cant conform to the the Emulex specification due to the implementation on linux.

Reply via email to