Re: [PATCH net-next 10/11] net: marvell: prestera: add storm control (rate limiter) implementation

[Ido Schimmel <idosch@idosch.org>]

On Fri, Jun 11, 2021 at 07:08:00PM +0200, Andrew Lunn wrote:
> On Fri, Jun 11, 2021 at 01:19:13PM +0000, Oleksandr Mazur wrote:
> > >>  On Wed, Jun 09, 2021 at 06:16:00PM +0300, Oleksandr Mazur wrote:
> > > Storm control (BUM) provides a mechanism to limit rate of ingress
> > > > port traffic (matched by type). Devlink port parameter API is used:
> > > > driver registers a set of per-port parameters that can be accessed to both
> > > > get/set per-port per-type rate limit.
> > > > Add new FW command - RATE_LIMIT_MODE_SET.
> > 
> > > Hi Oleksandr
> > 
> > > Just expanding on the two comments you already received about this.
> > 
> > > We often see people miss that switchdev is about. It is not about
> > > writing switch drivers. It is about writing network stack
> > > accelerators. You take a feature of the Linux network stack and you
> > > accelerate it by offloading it to the hardware. So look around the
> > > network stack and see how you configure it to perform rate limiting of
> > > broadcast traffic ingress. Once you have found a suitable mechanism,
> > > accelerate it via offloading.
> > 
> > > If you find Linux has no way to perform a feature the hardware could
> > > accelerate, you first need to add a pure software version of that
> > > feature to the network stack, and then add acceleration support for
> > > it.
> > 
> > 
> > Hello Andrew, Ido, Nikolay,
> > I appreciate your time and comments provided over this patchset, though i have a few questions to ask, if you don't mind:
> > 
> 
> > 1. Does it mean that in order to support storm control in switchdev
> > driver i need to implement software storm control in bridge driver,
> > and then using the switchdev attributes (notifiers) mechanism
> > offload the configuration itself to the HW?
> 
> Hi Oleksandr
> 
> Not necessarily. Is storm control anything more than ingress packet
> matching and rate limiting?
> 
> I'm not TC expert, but look for example at
> https://man7.org/linux/man-pages/man8/tc-police.8.html
> 
> and the example:
> 
> # tc qdisc add dev eth0 handle ffff: ingress
> # tc filter add dev eth0 parent ffff: u32 \
>                    match u32 0 0 \
>                    police rate 1mbit burst 100k
> 
> Replace the "match u32 0 0" with something which matches on broadcast
> frames.  Maybe "flower dst_mac ff:ff:ff:ff:ff:ff"
> 
> So there is a software solution. Now accelerate it.

Storm control also needs the ability to limit other types of flooded
traffic such unknown unicast and unregistered multicast packets. The
entity which classifies packets as such is the bridge, which happens
after the ingress hook.

I see two options to support storm control in Linux:

1. By adding support in the bridge itself as a new bridge slave option.
Something like:

# ip link set dev swp1 type bridge_slave \
	storm_control type { uuc | umc | bc} rate RATE mode { packet | byte }

I suspect this similar to more traditional implementations that users
might be used to and also maps nicely to hardware implementations

2. Teaching tc to call into the bridge to classify a packet. Not sure a
whole new classifier is needed for this. Maybe just extend flower with a
new key: dst_mac_type { uuc | umc }. I personally find this a bit weird,
but it is more flexible and allows to reuse existing actions

> 
> > 2. Is there any chance of keeping devlink solution untill the
> > discussed (storm control implemented in the bridge driver) mechanism
> > will be ready/implemented?
> 
> No. Please do it correctly from the beginning. No hacks.

+1