[ovs-dev] [PATCH v2 1/1] netdev-dpdk: fix RX jumbo for NICs not supporting scatter

Pablo Cascón pablo.cascon at netronome.com
Wed Apr 18 14:41:41 UTC 2018


On 13/04/18 19:45, Kevin Traynor wrote:
> On 04/13/2018 04:20 PM, Stokes, Ian wrote:
>>> Currently to RX jumbo packets fails for NICs not supporting scatter.
>>> Scatter is not strictly needed for jumbo support on RX. This change fixes
>>> the issue by only enabling scatter for NICs supporting it. Add a quirk for
>>> "igb" while the PMD is fixed to advertise scatter.
>>>
>> Thanks for the v2 Pablo.
>>
>> Adding Eelco and Kevin as they had some comments on the v1.
>>
>> FYI I'm investigating on the DPDK side to see how/when the flag should be set and used for igb and ixgbe as well as other drivers.
>>
>> https://dpdk.org/ml/archives/dev/2018-April/097056.html
>>
>>> Reported-by: Louis Peens<louis.peens at netronome.com>
>>> Signed-off-by: Pablo Cascón<pablo.cascon at netronome.com>
>>> Reviewed-by: Simon Horman<simon.horman at netronome.com>
>>> ---
>>>   lib/netdev-dpdk.c | 10 +++++++++-
>>>   1 file changed, 9 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c index ee39cbe..8f6a0a3
>>> 100644
>>> --- a/lib/netdev-dpdk.c
>>> +++ b/lib/netdev-dpdk.c
>>> @@ -694,11 +694,19 @@ dpdk_eth_dev_queue_setup(struct netdev_dpdk *dev,
>>> int n_rxq, int n_txq)
>>>       int diag = 0;
>>>       int i;
>>>       struct rte_eth_conf conf = port_conf;
>>> +    struct rte_eth_dev_info info;
>>>
>>>       /* For some NICs (e.g. Niantic), scatter_rx mode needs to be
>>> explicitly
>>>        * enabled. */
>>>       if (dev->mtu > ETHER_MTU) {
>>> -        conf.rxmode.enable_scatter = 1;
>>> +        rte_eth_dev_info_get(dev->port_id, &info);
>>> +        if (info.rx_offload_capa & DEV_RX_OFFLOAD_SCATTER) {
>>> +            conf.rxmode.enable_scatter = 1;
>>> +        } else if (!strcmp(info.driver_name, "igb")) {
>>> +            /* Quirk: as of DPDK 17.11.1 igb's PMD requires explicitly
>>> +               enabling scatter but fails to advertise it. */
> Can you check name for "nfp" and don't set enable_scatter? I don't think
> most of the PMDs used these new offload defines in DPDK17.11.

(sorry for the delay replying, was on PTO)

While it is technically possible to do that for the "nfp" it is not the 
preferred solution IMHO. The "nfp" PMD is doing the right thing (TM) by 
not advertising it supports scatter (it doesn't) and not requiring it 
for jumbo traffic. If we're to add a quirk it should be for the 
to-be-fixed PMD(s).

>> I'm not sure this is acceptable. I'm worried it sets a precedent for code for specific devices and could lead to further instances of this in the future.
>>
>> It could be argued the scatter approach up to now was specific to Niantic but it also worked for igb and i40e. I40e devices don’t require scatter but can handle it without issue if it is set.
>>
>> In the past this type of solution has been rejected as the preferred approach was to keep netdev-dpdk code generic as possible.

The principle makes sense in general to me. I think this should be a 
temporary exception.

>>
>> That’s why I suggest deferring the patch in OVS until the required changes are made in DPDK to satisfy all cases. 17.11.2 is targeted for May 19th. We could have a solution in place for then.

That would be fine when it comes to releases, not so fine for 
(potential) backports in distros and upstream consumers (interested on 
commits on between releases)

>>
>> I'm not trying to obstruct this but these cases do arise so interested to hear what others think?

+1

>>
>> Ian
>>
>>> +            conf.rxmode.enable_scatter = 1;
>>> +        }
>>>       }
>>>
>>>       conf.rxmode.hw_ip_checksum = (dev->hw_ol_features &
>>> --
>>> 2.7.4
>>>
>>> _______________________________________________
>>> dev mailing list
>>> dev at openvswitch.org
>>> https://mail.openvswitch.org/mailman/listinfo/ovs-dev



More information about the dev mailing list