[ovs-dev] [PATCH v2 6/8] netdev-dpdk: retry with queue action
Finn Christensen
fc at napatech.com
Fri Sep 15 19:55:27 UTC 2017
-----Original Message-----
From: Darrell Ball [mailto:dball at vmware.com]
Sent: 15. september 2017 20:09
To: Finn Christensen <fc at napatech.com>
Cc: dev at openvswitch.org
Subject: Re: [ovs-dev] [PATCH v2 6/8] netdev-dpdk: retry with queue action
On 9/15/17, 2:16 AM, "Finn Christensen" <fc at napatech.com> wrote:
-----Original Message-----
From: Darrell Ball [mailto:dball at vmware.com]
Sent: 14. september 2017 21:35
To: Finn Christensen <fc at napatech.com>
Cc: dev at openvswitch.org
Subject: Re: [ovs-dev] [PATCH v2 6/8] netdev-dpdk: retry with queue
action
On 9/14/17, 1:14 AM, "Finn Christensen" <fc at napatech.com> wrote:
-----Original Message-----
From: ovs-dev-bounces at openvswitch.org [mailto:ovs-dev-
bounces at openvswitch.org] On Behalf Of Darrell Ball
Sent: 13. september 2017 18:18
To: Simon Horman <simon.horman at netronome.com>
Cc: dev at openvswitch.org
Subject: Re: [ovs-dev] [PATCH v2 6/8] netdev-dpdk: retry with queue
action
On 9/13/17, 2:57 AM, "Simon Horman"
<simon.horman at netronome.com> wrote:
On Tue, Sep 12, 2017 at 08:36:19AM +0000, Darrell Ball wrote:
>
> On 9/10/17, 11:14 PM, "ovs-dev-bounces at openvswitch.org on
behalf of Yuanhan Liu" <ovs-dev-bounces at openvswitch.org on behalf
of
yliu at fridaylinux.org> wrote:
>
> On Fri, Sep 08, 2017 at 06:48:50PM +0200, Simon Horman
wrote:
> > On Tue, Sep 05, 2017 at 05:22:59PM +0800, Yuanhan Liu
wrote:
> > > From: Finn Christensen <fc at napatech.com>
> > >
> > > AFAIK, most (if not all) NICs (including Mellanox and Intel)
do
not
> > > support a pure MARK action. It's required to be used
together
with
> > > some other actions, like QUEUE.
> > >
> > > To workaround it, retry with a queue action when first try
failed.
> > >
> > > Moreover, some Intel's NIC (say XL710) needs the QUEUE
action set
> > > before the MARK action.
> > >
> > > Co-authored-by: Yuanhan Liu <yliu at fridaylinux.org>
> > > Signed-off-by: Finn Christensen <fc at napatech.com>
> > > Signed-off-by: Yuanhan Liu <yliu at fridaylinux.org>
> >
> > This feels a bit like the tail wagging the dog.
> > Is this the lowest level at which it makes sense to implement
> > this logic?
> >
> > If so then I wonder if some sort of probing would be in order
> > to avoid the cost of trying to add the flow twice to hardware
> > where the queue is required.
>
> Do you mean something like rte_flow capability query, like
whether
> a queue action is needed for a mark action? If so, yes, I do
think
> we miss an interface like this.
>
> Note that even in this solution, the flow won't be created
twice
> to the hardware, because the first try would be failed.
>
> [Darrell]
>
> Having an api to quey capability and avoid the first try to
HW
would be nice, but there are dependencies
> on RTE, drivers etc and I don’t know definitive the api
would
be.
>
> Also, as nics are added this capability needs to be done
and
state needs to be kept in all cases.
>
> It is an enhancement and if done should be reliable.
Agreed. Though I was more thinking of probing the hardware
rather
than
having a capability API - I expect this would remove several of the
dependencies you describe above.
[Darrell] I have been pondering the probing option as well. It is
certainly a
valid option; we use it in other cases such as datapath probing. One of
the
aspects that worries me here is maintaining the correct per interface
(essentially; although the attribute is per nic) state across various
events
such as new ports being added, vswitchd restarts, races with flow
creation.
It would be non-trivial I guess and probably appropriate for the next
patch
series, if done.
In this case, we have what seems like a clear distinction b/w
Napatech
which does not need the queue action workaround and everything
else,
which does.
Besides the non-Napatech behavior, which is worrisome,
maintaining the
difference for flow handling under the covers is concerning.
I wonder if we should be upfront as possible here and just have a
dpdk
interface configuration – maybe something like “supports native
HWOL
mark action” since the better behavior is the exception?
The interface config would be more robust than probing.
This would need documentation, of course.
[Finn]
I think the rte queue action should never be used here when using
partial HWOL. Not the way OVS handles multi queues today.
Maybe a "default queue" could be used in the dpdk PMDs when no
queue is specified in rte flow?
[Darrell] This is the Napatech case, where no queue action is needed;
you
are suggesting programming a default queue in this case.
I don’t follow how this would be helpful/desired?
[Finn]
I was trying to make my view on this, not particularly arguing for the
Napatech
Case.
Here is what I was thinking:
Taking the case of this partial HWOL, then we are trying to offload the
flow
classification to HW, like "pre-classify and mark". Then this mark is used
to
accelerate OVS while finding the actions to execute. Since we do leave all
processing of the actions to OVS, there is no way for the partial HWOL to
know, at rte flow creation time, where to send the pre-classified packets
(which is strictly needed when seen from a DPDK rte flow point of view).
When multiple queues are specified in OVS, a hash splitting mechanism
is used in the nic to support RSS. Then the nic is responsible for selecting
the
right queue according to the configured algorithm for RSS.
OVS only needs to know how many queues to service per port. No
knowledge
about the association b/w flow <-> rxq is used in OVS today.
When looking at full HWOL, all flow actions will have to be interpreted,
supported
and programmed to NIC using rte flow - send directly to target port and
we
do not have this issue.
Have I missed something?
[Darrell]
I understand you points about the full offload you are working on.
I just did not follow this one comment:
“Maybe a "default queue" could be used in the dpdk PMDs when no
queue is specified in rte flow?”
which I thought is related to partial offloads used here and the receive
queue action we are discussing?; I am concerned that I missed your point
about this specific comment ?
[Finn]
Sorry for my messy descriptions, and yes the "default queue" idea was for the
partial HWOL case. I just think if the queue could be set to 0 if 1 rxq used and let
the hash algorithm set the queue if RSS is used, then we do not have to add QUEUE
to the rte flow create.
However, it was just an idea and I don't know if it can be done reliably in all DPDK PMDs.
Essentially, this is a mismatch between the rte flow impl
functionality in
PMD and the needs in OVS for flow classification, in a partial HWOL
setup.
[Darrell] Maybe this has been explored, but were additional
workaround
actions, like RTE_FLOW_ACTION_TYPE_RSS considered ?
I guess we could make this essentially a NOP, but if mark
action
could ride along with it, then that would be good.
[Finn]
No. We have not explored that.
I'm not sure I understand how you would use
RTE_FLOW_ACTION_TYPE_RSS
In this situation.
[Darrell]
So there are two points here:
1/ ‘IF’ RTE_FLOW_ACTION_TYPE_RSS can used in lieu of
RTE_FLOW_ACTION_TYPE_QUEUE combined with
RTE_FLOW_ACTION_TYPE_MARK, as need for the workaround for most
nics, then this would be better for distributing packets across PMDs for the
masked flows relevant here. This would help mitigate one of the concerns.
We could use the RSS we use globally, essentially making the distribution a
NOP compared with what we normally do today with RSS….
2/ Assuming ‘1’, we might even use this in all cases as a first try for all nics,
which if it could be done solves the queue action workaround
capability/probing thing?
[Finn] Thanks, I see what you mean. It makes very good sense if it is supported
on the nic DPDK PMDs.
However, we would not mind Darells proposal, it makes sense since
most nic PMDs unfortunately needs this and it is only relevant for
partial
HWOL. We are mostly concerned with full HWOL and therefore see
partial
HWOL as a failover when full HWOL could not be handled in nic, if
enabled.
I think anyways we need documentation describing the difference
b/w
nics in the dpdk documentation (howto part).
Assuming no such enhancement is appropriate at this time I
would
still like to ask if this is the best place for this hardware-specific
code?
[Darrell]
For OVS, the netdev-dpdk layer is the lowest layer.
This kind of workaround is hard to hide, since we are messing with
the
rxq, so I think OVS needs to know that it is in effect anyways. An
alternative is to supply a mark and an ‘optional queue’ and let the
driver
decide if the queue is needed and report back whether it was. This
would
be hard to do across various drivers. Supporting in the rte layer would
require both rte and driver support, so even more support.
> A separate comment is we need to document which nics
need
the queue action.
>
> Also, I think we should check errno in the present code.
_______________________________________________
dev mailing list
dev at openvswitch.org
https://urldefense.proofpoint.com/v2/url?u=https-
3A__mail.openvswitch.org_mailman_listinfo_ovs-
2Ddev&d=DwIGaQ&c=uilaK90D4TOVoH58JNXRgQ&r=BVhFA09CGX7JQ5Ih-
uZnsw&m=IHypHavCy0AKjNxqOMyc4w3ILyC-
BuwkB8fuVvQUA3k&s=deJQWP9KI22Xp46tEoZ6o6Emitr3Bhfd7iSMxNpude
g&e=
Disclaimer: This email and any files transmitted with it may contain
confidential information intended for the addressee(s) only. The
information is not to be surrendered or copied to unauthorized
persons. If
you have received this communication in error, please notify the
sender
immediately and delete this e-mail from your system.
Disclaimer: This email and any files transmitted with it may contain
confidential information intended for the addressee(s) only. The
information is not to be surrendered or copied to unauthorized persons. If
you have received this communication in error, please notify the sender
immediately and delete this e-mail from your system.
More information about the dev
mailing list