Re: [RFC PATCH net-next 0/5] Replay and offload host VLAN entries in DSA

On 14/02/2022 12:42, Vladimir Oltean wrote:

On Mon, Feb 14, 2022 at 12:27:57PM +0200, Nikolay Aleksandrov wrote:

quoted

On 14/02/2022 12:07, Vladimir Oltean wrote:

quoted

On Mon, Feb 14, 2022 at 11:59:02AM +0200, Ido Schimmel wrote:

quoted

Sounds good to me as well. I assume it means patches #1 and #2 will be
changed to make use of this flag and patch #3 will be dropped?

Not quite. Patches 1 and 2 will be kept as-is, since fundamentally,
their goal is to eliminate a useless SWITCHDEV_PORT_OBJ_ADD event when
really nothing has changed (flags == old_flags, no brentry was created).

I don't think that's needed, a two-line change like
"vlan_already_exists == true && old_flags == flags then do nothing"
would do the trick in DSA for all drivers and avoid the churn of these patches.
It will also keep the order of the events consistent with (most of) the rest
of the bridge. I'd prefer the simpler change which avoids config reverts than
these two patches.

I understand you prefer the simpler change which avoids reverting the
struct net_bridge_vlan on error, but "vlan_already_exists == true &&
old_flags == flags then do nothing" is not possible in DSA, because DSA
does not know "old_flags" unless we also pass that via
struct switchdev_obj_port_vlan. If we do that, I don't have much of an
objection, but it still seems cleaner to me if the bridge didn't notify
switchdev at all when it has nothing to say.

Or where do you mean to place the two-line change?

You mention a couple of times in both patches that you'd like to add dsa
vlan refcounting infra similar to dsa's host fdb and mdb. So I assumed that involves
keeping track of vlan entries in dsa? If so, then I thought you'd record the old flags there.

Alternatively I don't mind passing old flags if you don't intend on keeping
vlan information in dsa, it's uglier but it will avoid the reverts which will
also avoid additional notifications when these cases are hit. It makes sense
to have both old flags and new flags if the switchdev checks are done pre-config
change so it can veto any transitions it can't handle for some reason.

A third option is to do the flags check in the bridge and avoid doing the
switchdev call (e.g. in br_switchdev_port_vlan_ calls), that should avoid
the reverts as well.

If you intend to add vlan refcounting in dsa, then I'd go with just keeping track
of the flags, you'll have vlan state anyway, otherwise it's up to you. I'm ok
with both options for old flags. 

Cheers,
 Nik