Re: [PATCH net-next 0/4] net: switchdev: Avoid sleep in atomic with DSA

[Jiri Pirko <jiri@resnulli.us>]

Tue, Jan 10, 2017 at 05:08:52PM CET, idosch@idosch.org wrote:
>On Tue, Jan 10, 2017 at 03:18:07PM +0100, Jiri Pirko wrote:
>> Tue, Jan 10, 2017 at 02:25:06PM CET, idosch@idosch.org wrote:
>> >On Tue, Jan 10, 2017 at 01:08:46PM +0100, Jiri Pirko wrote:
>> >> Mon, Jan 09, 2017 at 10:14:36PM CET, idosch@idosch.org wrote:
>> >> >On Mon, Jan 09, 2017 at 12:56:48PM -0800, Florian Fainelli wrote:
>> >> >> On 01/09/2017 12:48 PM, Ido Schimmel wrote:
>> >> >> > Hi Florian,
>> >> >> > 
>> >> >> > On Mon, Jan 09, 2017 at 11:44:59AM -0800, Florian Fainelli wrote:
>> >> >> >> Hi all,
>> >> >> >>
>> >> >> >> This patch series is to resolve a sleeping function called in atomic context
>> >> >> >> debug splat that we observe with DSA.
>> >> >> >>
>> >> >> >> Let me know what you think, I was also wondering if we should just always
>> >> >> >> make switchdev_port_vlan_fill() set SWITCHDEV_F_DEFER, but was afraid this
>> >> >> >> could cause invalid contexts to be used for rocker, mlxsw, i40e etc.
>> >> >> > 
>> >> >> > Isn't this a bit of overkill? All the drivers you mention fill the VLAN
>> >> >> > dump from their cache and don't require sleeping. Even b53 that you
>> >> >> > mention in the last patch does that, but reads the PVID from the device,
>> >> >> > which entails taking a mutex.
>> >> >> 
>> >> >> Correct.
>> >> >> 
>> >> >> > 
>> >> >> > Can't you just cache the PVID as well? I think this will solve your
>> >> >> > problem. Didn't look too much into the b53 code, so maybe I'm missing
>> >> >> > something. Seems that mv88e6xxx has a similar problem.
>> >> >> 
>> >> >> I suppose we could indeed cache the PVID for b53, but for mv88e6xxx it
>> >> >> seems like we need to perform a bunch of VTU operations, and those
>> >> >> access HW registers, Andrew, Vivien, how do you want to solve that, do
>> >> >> we want to introduce a general VLAN cache somewhere in switchdev/DSA/driver?
>> >> >
>> >> >Truth be told, I don't quite understand why switchdev infra even tries
>> >> >to dump the VLANs from the device. Like, in which situations is this
>> >> >going to be different from what the software bridge reports? Sure, you
>> >> >can set the VLAN filters with SELF and skip the software bridge, but how
>> >> >does that make sense in a model where you want to reflect the software
>> >> >datapath?
>> >> 
>> >> But the vlans added by rtnl_bridge_setlink & SELF are not tracked by the
>> >> bridge and therefore driver needs to dump them. You would have to pass
>> >> some flag down to driver when adding SWITCHDEV_OBJ_ID_PORT_VLAN
>> >> indicating the need to track the vlan and dump it. Right?
>> >
>> >Right, but back to my question - what's the use case for the SELF flag
>> >in the switchdev model? Why would I configure a VLAN filter in the
>> >hardware but not in the software bridge? The whole point is reflecting
>> >the software bridge to the hardware.
>> 
>> I agree. For the bridge-switchdev usecase, I don't see a reason to use
>> SELF for vlans as well. Do you suggest to simply remove this possibility?
>
>Yes. This would also solve Florian's problem.

I agree, for switchdev usecase it makes no sense to implement this. I
vote for removing that.


>
>Florian, what do you think?