Date: Mon, 10 Jul 2023 14:21:24 -0400
From: Eric Garver <eric@...ver.life>
To: Ilya Maximets <i.maximets@....org>
Cc: Jakub Kicinski <kuba@...nel.org>, Aaron Conole <aconole@...hat.com>,
	netdev@...r.kernel.org, dev@...nvswitch.org,
	Paolo Abeni <pabeni@...hat.com>, Eric Dumazet <edumazet@...gle.com>,
	"David S. Miller" <davem@...emloft.net>,
	Adrian Moreno <amorenoz@...hat.com>,
	Eelco Chaudron <echaudro@...hat.com>
Subject: Re: [ovs-dev] [PATCH net-next 2/2] net: openvswitch: add drop action

On Mon, Jul 10, 2023 at 06:51:19PM +0200, Ilya Maximets wrote:
> On 7/8/23 00:06, Jakub Kicinski wrote:
> > On Fri, 7 Jul 2023 18:04:36 +0200 Ilya Maximets wrote:
> >>>> That already exists, right? Johannes added it in the last release for WiFi.  
> >>>
> >>> I'm not sure.  The SKB_DROP_REASON_SUBSYS_MAC80211_UNUSABLE behaves similarly
> >>> to that on a surface.  However, looking closer, any value that can be passed
> >>> into ieee80211_rx_handlers_result() and ends up in the kfree_skb_reason() is
> >>> kind of defined in net/mac80211/drop.h, unless I'm missing something (very
> >>> possible, because I don't really know wifi code).
> >>>
> >>> The difference, I guess, is that for openvswitch values will be provided by
> >>> the userpsace application via netlink interface.  It'll be just a number not
> >>> defined anywhere in the kernel.  Only the subsystem itself will be defined
> >>> in order to occupy the range.  Garbage in, same garbage out, from the kernel's
> >>> perspective.  
> >>
> >> To be clear, I think, not defining them in this particular case is better.
> >> Definition of every reason that userspace can come up with will add extra
> >> uAPI maintenance cost/issues with no practical benefits.  Values are not
> >> going to be used for anything outside reporting a drop reason and subsystem
> >> offset is not part of uAPI anyway.
> > 
> > Ah, I see. No, please don't stuff user space defined values into 
> > the drop reason. The reasons are for debugging the kernel stack 
> > itself. IOW it'd be abuse not reuse.
> 
> Makes sense.  I wasn't sure that's a good solution from a kernel perspective
> either.  It's better than defining all these reasons, IMO, but it's not good
> enough to be considered acceptable, I agree.
> 
> How about we define just 2 reasons, e.g. OVS_DROP_REASON_EXPLICIT_ACTION and
> OVS_DROP_REASON_EXPLICIT_ACTION_WITH_ERROR (exact names can be different) ?
> One for an explicit drop action with a zero argument and one for an explicit
> drop with non-zero argument.
> 
> The exact reason for the error can be retrieved by other means, i.e by looking
> at the datapath flow dump or OVS logs/traces.
> 
> This way we can give a user who is catching packet drop traces a signal that
> there was something wrong with an OVS flow and they can look up exact details
> from the userspace / flow dump.
> 
> The point being, most of the flows will have a zero as a drop action argument,
> i.e. a regular explicit packet drop.  It will be hard to figure out which flow
> exactly we're hitting without looking at the full flow dump.  And if the value
> is non-zero, then it should be immediately obvious which flow is to blame from
> the dump, as we should not have a lot of such flows.
> 
> This would still allow us to avoid a maintenance burden of defining every case,
> which are fairly meaningless for the kernel itself, while having 99% of the
> information we may need.
> 
> Jakub, do you think this will be acceptable?
> 
> Eric, Adrian, Aaron, do you see any problems with such implementation?

I see no problems. I'm content with this approach.

> P.S. There is a plan to add more drop reasons for other places in openvswitch
>      module to catch more regular types of drops like memory issues or upcall
>      failures.  So, the drop reason subsystem can be extended later.
>      The explicit drop action is a bit of an odd case here.
> 
> Best regards, Ilya Maximets.
> 

Powered by blists - more mailing lists