Re: [PATCH 0/7] net: dsa: mv88e6xxx: features to handle network storms

[Florian Fainelli <f.fainelli@gmail.com>]

On 9/12/19 9:46 AM, Robert Beckett wrote:
> On Thu, 2019-09-12 at 09:25 -0700, Florian Fainelli wrote:
>> On 9/12/19 2:03 AM, Ido Schimmel wrote:
>>> On Wed, Sep 11, 2019 at 12:49:03PM +0100, Robert Beckett wrote:
>>>> On Wed, 2019-09-11 at 11:21 +0000, Ido Schimmel wrote:
>>>>> On Tue, Sep 10, 2019 at 09:49:46AM -0700, Florian Fainelli
>>>>> wrote:
>>>>>> +Ido, Jiri,
>>>>>>
>>>>>> On 9/10/19 8:41 AM, Robert Beckett wrote:
>>>>>>> This patch-set adds support for some features of the
>>>>>>> Marvell
>>>>>>> switch
>>>>>>> chips that can be used to handle packet storms.
>>>>>>>
>>>>>>> The rationale for this was a setup that requires the
>>>>>>> ability to
>>>>>>> receive
>>>>>>> traffic from one port, while a packet storm is occuring on
>>>>>>> another port
>>>>>>> (via an external switch with a deliberate loop). This is
>>>>>>> needed
>>>>>>> to
>>>>>>> ensure vital data delivery from a specific port, while
>>>>>>> mitigating
>>>>>>> any
>>>>>>> loops or DoS that a user may introduce on another port
>>>>>>> (can't
>>>>>>> guarantee
>>>>>>> sensible users).
>>>>>>
>>>>>> The use case is reasonable, but the implementation is not
>>>>>> really.
>>>>>> You
>>>>>> are using Device Tree which is meant to describe hardware as
>>>>>> a
>>>>>> policy
>>>>>> holder for setting up queue priorities and likewise for queue
>>>>>> scheduling.
>>>>>>
>>>>>> The tool that should be used for that purpose is tc and
>>>>>> possibly an
>>>>>> appropriately offloaded queue scheduler in order to map the
>>>>>> desired
>>>>>> scheduling class to what the hardware supports.
>>>>>>
>>>>>> Jiri, Ido, how do you guys support this with mlxsw?
>>>>>
>>>>> Hi Florian,
>>>>>
>>>>> Are you referring to policing traffic towards the CPU using a
>>>>> policer
>>>>> on
>>>>> the egress of the CPU port? At least that's what I understand
>>>>> from
>>>>> the
>>>>> description of patch 6 below.
>>>>>
>>>>> If so, mlxsw sets policers for different traffic types during
>>>>> its
>>>>> initialization sequence. These policers are not exposed to the
>>>>> user
>>>>> nor
>>>>> configurable. While the default settings are good for most
>>>>> users, we
>>>>> do
>>>>> want to allow users to change these and expose current
>>>>> settings.
>>>>>
>>>>> I agree that tc seems like the right choice, but the question
>>>>> is
>>>>> where
>>>>> are we going to install the filters?
>>>>>
>>>>
>>>> Before I go too far down the rabbit hole of tc traffic shaping,
>>>> maybe
>>>> it would be good to explain in more detail the problem I am
>>>> trying to
>>>> solve.
>>>>
>>>> We have a setup as follows:
>>>>
>>>> Marvell 88E6240 switch chip, accepting traffic from 4 ports. Port
>>>> 1
>>>> (P1) is critical priority, no dropped packets allowed, all others
>>>> can
>>>> be best effort.
>>>>
>>>> CPU port of swtich chip is connected via phy to phy of intel i210
>>>> (igb
>>>> driver).
>>>>
>>>> i210 is connected via pcie switch to imx6.
>>>>
>>>> When too many small packets attempt to be delivered to CPU port
>>>> (e.g.
>>>> during broadcast flood) we saw dropped packets.
>>>>
>>>> The packets were being received by i210 in to rx descriptor
>>>> buffer
>>>> fine, but the CPU could not keep up with the load. We saw
>>>> rx_fifo_errors increasing rapidly and ksoftirqd at ~100% CPU.
>>>>
>>>>
>>>> With this in mind, I am wondering whether any amount of tc
>>>> traffic
>>>> shaping would help? Would tc shaping require that the packet
>>>> reception
>>>> manages to keep up before it can enact its policies? Does the
>>>> infrastructure have accelerator offload hooks to be able to apply
>>>> it
>>>> via HW? I dont see how it would be able to inspect the packets to
>>>> apply
>>>> filtering if they were dropped due to rx descriptor exhaustion.
>>>> (please
>>>> bear with me with the basic questions, I am not familiar with
>>>> this part
>>>> of the stack).
>>>>
>>>> Assuming that tc is still the way to go, after a brief look in to
>>>> the
>>>> man pages and the documentation at largc.org, it seems like it
>>>> would
>>>> need to use the ingress qdisc, with some sort of system to
>>>> segregate
>>>> and priortise based on ingress port. Is this possible?
>>>
>>> Hi Robert,
>>>
>>> As I see it, you have two problems here:
>>>
>>> 1. Classification: Based on ingress port in your case
>>>
>>> 2. Scheduling: How to schedule between the different transmission
>>> queues
>>>
>>> Where the port from which the packets should egress is the CPU
>>> port,
>>> before they cross the PCI towards the imx6.
>>>
>>> Both of these issues can be solved by tc. The main problem is that
>>> today
>>> we do not have a netdev to represent the CPU port and therefore
>>> can't
>>> use existing infra like tc. I believe we need to create one.
>>> Besides
>>> scheduling, we can also use it to permit/deny certain traffic from
>>> reaching the CPU and perform policing.
>>
>> We do not necessarily have to create a CPU netdev, we can overlay
>> netdev
>> operations onto the DSA master interface (fec in that case), and
>> whenever you configure the DSA master interface, we also call back
>> into
>> the switch side for the CPU port. This is not necessarily the
>> cleanest
>> way to do things, but that is how we support ethtool operations (and
>> some netdev operations incidentally), and it works
> 
> After reading up on tc, I am not sure how this would work given the
> semantics of the tool currently.
> 
> My initial thought was to model the switch's 4 output queues using an
> mqprio qdisc for the CPU port, and then use either iptables's classify
> module on the input ports to set which queue it egresses from on the
> CPU port, or use vlan tagging with id 0 and priority set. (with the
> many detail of how to implement them still left to discover).
> 
> However, it looks like the mqprio qdisc could only be used for egress,
> so without a netdev representing the CPU port, I dont know how it could
> be used.

If you are looking at mapping your DSA master/CPU port egress queues to
actual switch egress queues, you can look at what bcm_sf2.c and
bcmsysport.c do and read the commit messages that introduced the mapping
functionality for background on why this was done. In a nutshell, the
hardware has the ability to back pressure the Ethernet MAC behind the
CPU port in order to automatically rate limit the egress out of the
switch. So for instance, if your CPU tries to send 1Gb/sec of traffic to
a port that is linked to a link partner at 100Mbits/sec, there is out of
band information between the switch and the Ethernet DMA of the CPU port
to pace the TX completion interrupt rate to match 100Mbits/sec.

This is going to be different for you here obviously because the
hardware has not been specifically designed for that, so you do need to
rely on more standard constructs, like actual egress QoS on both ends.

> 
> Another thing I thought of using was just to use iptable's TOS module
> to set the minimal delay bit and rely on default behaviours, but Ive
> yet to find anything in the Marvell manual that indicates it could set
> that bit on all frames entering a port.
> 
> Another option might be to use vlans with their priority bits being
> used to steer to output queues, but I really dont want to introduce
> more virtual interfaces in to the setup, and I cant see how to
> configure an enforce default vlan tag with id 0 and priority bits set
> via linux userland tools.
> 
> 
> It does look like tc would be quite nice for configuring the egress
> rate limiting assuming we a netdev to target with the rate controls of
> the qdisc.
> 
> 
> So far, this seems like I am trying to shoe horn this stuff in to tc.
> It seems like tc is meant to configure how the ip stack  configures
> flow within the stack, whereas in a switch chip, the packets go nowhere
> near the CPUs kernel ip stack. I cant help thinking that it would be
> good have a specific utility for configuring switches that operates on
> the port level for manage flow within the chip, or maybe simple sysfs
> attributes to set the ports priority.

I am not looking at tc the same way you are doing, tc is just the tool
to configure all QoS/ingress/egress related operations on a network
device. Whether that network device can offload some of the TC
operations or not is where things get interesting.

TC has ingress filtering support, which is what you could use for
offloading broadcast storms, I would imagine that the following should
be possible to be offloaded (this is not a working command but you get
the idea):

tc qdisc add dev sw0p0 handle ffff: ingress
tc filter add dev sw0p0 parent ffff: protocol ip prio 1 u32 match ether
src 0xfffffffffffff police rate 100k burst 10k skip_sw

something along those lines is how I would implement ingress rate
limiting leveraging what the switch could do. This might mean adding
support for offloading specific TC filters, Jiri and Ido can certainly
suggest a cleverer way of achieving that same functionality.
-- 
Florian