On 16 Mar 2023, at 11:48, Ilya Maximets wrote:
> On 3/16/23 10:15, Eelco Chaudron wrote:
>> On 1 Mar 2023, at 8:22, Chris Mi wrote:
>>
>>> Iterate each registered offload API. It's not a problem for today
>>> since we only have one implementation.
>>
>> See some comments inline below.
>>
>> //Eelco
>>
>>> Signed-off-by: Chris Mi <c...@nvidia.com>
>>> Reviewed-by: Roi Dayan <r...@nvidia.com>
>>> ---
>>> lib/netdev-offload.c | 35 +++++++++++++++++++++++++++++++++++
>>> lib/netdev-offload.h | 5 +++++
>>> 2 files changed, 40 insertions(+)
>>>
>>> diff --git a/lib/netdev-offload.c b/lib/netdev-offload.c
>>> index 4592262bd..d6e3e4946 100644
>>> --- a/lib/netdev-offload.c
>>> +++ b/lib/netdev-offload.c
>>> @@ -256,6 +256,41 @@ meter_offload_del(ofproto_meter_id meter_id, struct
>>> ofputil_meter_stats *stats)
>>> return 0;
>>> }
>>>
>>> +int
>>> +netdev_offload_recv(struct dpif_upcall *upcall, struct ofpbuf *buf,
>>> + uint32_t handler_id)
>>> +{
>>> + struct netdev_registered_flow_api *rfa;
>>> + int ret = EAGAIN;
>>> +
>>> + CMAP_FOR_EACH (rfa, cmap_node, &netdev_flow_apis) {
>>> + if (rfa->flow_api->recv) {
>>> + ret = rfa->flow_api->recv(upcall, buf, handler_id);
>>> + if (ret) {
>>> + VLOG_DBG_RL(&rl, "Failed to receive offload packet, %s, "
>>> + "type: %s", ovs_strerror(ret),
>>> + rfa->flow_api->type);
>>> + } else {
>>> + return ret;
>>
>> So how would this work if two or more offloads support this API?
>>
>> Only if the first one returns an error (if no packets are waiting, what
>> would the error be, we might want to exclude the debug message?), the second
>> one is tried.
>>
>> Also what about one starving the other?
>
> I guess, it was my suggestion to keep it simple for now and
> ignore the possible starvation. :)
>
> I'm not really sure if rte_flow will actually have API like
> this in a near future.
>
> We could introduce a different iteration way here when necessary.
> i.e. have a sequence number and call recv() first on APIs that
> have older value. But right now it's probably not worth it.
>
> What do you think?
I’m fine for now, guess it sounded odd when we call the potential rte_flow API
in threads explicitly created for dpif-netlink.
> Not logging on EAGAIN or something like that is a good idea though.
>
> Best regards, Ilya Maximets.
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
