Date:   Mon, 16 May 2022 02:57:50 +0200
From:   Pablo Neira Ayuso <pablo@...filter.org>
To:     Felix Fietkau <nbd@....name>
Cc:     netfilter-devel@...r.kernel.org, netdev@...r.kernel.org,
        Jo-Philipp Wich <jo@...n.io>
Subject: Re: [RFC] netfilter: nf_tables: ignore errors on flowtable device hw
 offload setup

On Fri, May 13, 2022 at 11:09:51AM +0200, Felix Fietkau wrote:
> 
> On 13.05.22 10:15, Pablo Neira Ayuso wrote:
> > On Fri, May 13, 2022 at 10:03:13AM +0200, Felix Fietkau wrote:
> > > 
> > > On 13.05.22 09:49, Pablo Neira Ayuso wrote:
> > > > Hi,
> > > > > On Tue, May 10, 2022 at 10:27:39PM +0200, Felix Fietkau wrote:
> > > > > In many cases, it's not easily possible for user space to know, which
> > > > > devices properly support hardware offload.
> > > > > Then, it is a matter of extending the netlink interface to
> > > expose this
> > > > feature? Probably add a FLOW_BLOCK_PROBE or similar which allow to
> > > > consult if this feature is available?
> > > > > > Even if a device supports hardware flow offload, it is not
> > > > > guaranteed that it will actually be able to handle the flows for
> > > > > which hardware offload is requested.
> > > > > When might this happen?
> > > 
> > > I think there are many possible reasons: The flow might be using features
> > > not supported by the offload driver. Maybe it doesn't have any space left in
> > > the offload table. I'm sure there are many other possible reasons it could
> > > fail.
> > 
> > This fallback to software flowtable path for partial scenarios already
> > exists.
> I know. All I meant was to point out that hardware offload is not guaranteed
> in one place, so I don't think bailing out with an error because flow block
> bind didn't work for one of the flowtable devices is justified.
> 
> > > > > Ignoring errors on the FLOW_BLOCK_BIND makes it a lot easier to set up
> > > > > configurations that use hardware offload where possible and gracefully
> > > > > fall back to software offload for everything else.
> > > > > I understand this might be useful from userspace perspective,
> > > because
> > > > forcing the user to re-try is silly.
> > > > > However, on the other hand, the user should have some way to
> > > know from
> > > > the control plane that the feature (hardware offload) that they
> > > > request is not available for their setup.
> > > 
> > > In my opinion, most users of this API probably don't care and just want to
> > > have offload on a best effort basis.
> > 
> > OK, but if the setup does not support hardware offload at all, why
> > should the control plane accept this? I think user should know in
> > first place that no one single flow is going to be offloaded to
> > hardware.
>
> It makes for a much cleaner configuration if you can just create a single
> hw-offload enabled flowtable containing multiple devices, some of which
> support hardware offload and some of which don't.

This scenario mixing devices that support hw offload and which does
not support hw offload make sense.

> > > Assuming that is the case, wouldn't it be better if we simply have
> > > an API that indicates, which flowtable members hardware offload was
> > > actually enabled for?
> > 
> > What are you proposing?
> > 
> > I think it would be good to expose through netlink interface what the
> > device can actually do according to the existing supported flowtable
> > software datapath features.
> In addition to the NFTA_FLOWTABLE_HOOK_DEVS array, the netlink API could
> also return another array, e.g. NFTA_FLOWTABLE_HOOK_OFFLOAD_DEVS which
> indicates devices for which hw offload is enabled.
> 
> What I really don't like about the current state of the flowtable offload
> API is the (in my opinion completely unnecessary) complexity that is
> required for the simple use case of enabling hw/sw flow offloading on a best
> effort basis for all devices.
> What I like even less is the number of implementation details that it has to
> consider.
> 
> For example: Let's assume we have a machine with several devices, some of
> which support hw offload, some of which don't. We have a mix of VLANs and
> bridges in there as well, maybe even PPPoE.
> Now the admin of that machine wants to enable best-effort hardware +
> software flow offloading for that configuration.
> Now he (or a piece of user space software dealing with the config) has to do
> these things:
> - figure out which devices could support hw offload, create a separate flow
> table for them
> - be aware of which of these devices are actually used by looking at the
> stack of bridges, vlans, dsa devices, etc.
> - if an error occurs, test them individually just to see which one actually
> failed and leave it out of the flowtable
> - for sw offload be aware that there is limited support for offloading decap
> of vlans/pppoe, count the number of decaps and figure out the right input
> device to add based on the behavior of nft_dev_path_info, so that the
> 'indev' it selects matches the device you put in the flow table.
> 
> So I'm asking you: Am I getting any of this completely wrong? Do you
> consider it to be a reasonable trade-off to force the admin (or intermediate
> user space layer) to jump through these hoops for such a simple use case,
> just because somebody might want more fine grained control?
> 
> I consider this patch to be a first step towards making simple use cases
> easier to configure. I'd also be fine with adding a flag to make the
> fallback behavior opt-in, even though I think it would make a much better
> default.
> 
> Eventually I'd also like to add a flag that makes it unnecessary to even
> specify the devices in the flow table by making the code auto-create hooks
> for devices with active flows, just like I did in my xtables target.
> You correctly pointed out to me in the past that this comes at the cost of a
> few packets delay before offloading kicks in, but I'm still wondering: who
> actually cares about that?
> 
> If I'm completely off-base with this, please let me know. I'm simply trying
> to make sense of all of this...

Maybe only fail if _none_ of the selected devices support for hardware
offload, ie. instead of silently accepting all devices, count of the
number of devices for which a block has been set up, if it is == 0
then bail out with EOPNOTSUPP.

Powered by blists - more mailing lists