On 16 Mar 2023, at 13:02, Ilya Maximets wrote:

> On 3/16/23 12:49, Eelco Chaudron wrote:
>>
>>
>> On 16 Mar 2023, at 12:00, Ilya Maximets wrote:
>>
>>> On 3/16/23 10:16, Eelco Chaudron wrote:
>>>> On 1 Mar 2023, at 8:22, Chris Mi wrote:
>>>>
>>>>> In thread handler 0, add netdev offload recv in normal recv upcalls.
>>>>> To avoid starvation, introduce a flag to alternate the order of
>>>>> receiving normal upcalls and offload upcalls based on that flag.
>>>>>
>>>>> Add similar change for recv_wait.
>>>>
>>>> See some comments inline below, and some architecture input from Ilya is 
>>>> needed...
>>>>
>>>> //Eelco
>>>>
>>>>
>>>>> Signed-off-by: Chris Mi <[email protected]>
>>>>> Reviewed-by: Roi Dayan <[email protected]>
>>>>> ---
>>>>>  lib/dpif-netlink.c            | 42 ++++++++++++++++++++++++++++++-----
>>>>>  ofproto/ofproto-dpif-upcall.c | 16 +++++++++----
>>>>>  2 files changed, 48 insertions(+), 10 deletions(-)
>>>
>>> <snip>
>>>
>>>>> +static int
>>>>> +dpif_netlink_recv(struct dpif *dpif_, uint32_t handler_id,
>>>>> +                  struct dpif_upcall *upcall, struct ofpbuf *buf)
>>>>> +{
>>>>> +    struct dpif_netlink *dpif = dpif_netlink_cast(dpif_);
>>>>> +    struct dpif_handler *handler;
>>>>> +    int error;
>>>>> +
>>>>> +    fat_rwlock_rdlock(&dpif->upcall_lock);
>>>>> +    handler = &dpif->handlers[handler_id];
>>>>> +    if (handler->recv_offload_first) {
>>>>> +        error = netdev_offload_recv(upcall, buf, handler_id);
>>>>
>>>> Here the abstraction blurs a bit, as netdev_offload_recv() would try to 
>>>> call implementations, tc and rte_flow?
>>>> See also earlier comment in netdev_offload_recv()
>>>>
>>>> Also what about for example rte_flow being handled now in the dpif-netdev 
>>>> class?
>>>> In general, OVS-DPDK does not have any upcall threads, so if OVS-DPDK (or 
>>>> any new offload) decides to use any of these mechanisms it will fail.
>>>>
>>>> Ilya do you have any feedback?
>>>
>>> IIRC, handlers are getting created.  But we need to implement recv()
>>> and recv_wait() callbacks for dpif-netdev, that will just call to
>>> netdev_offload_recv[_wait]().
>>
>> I’m ok for now (see other email), just to clarify that the handler threads 
>> get created out of dpif-netlink.c.
>> So in theory now we would be calling netdev_offload_recv[_wait]() in the 
>> dpif-netlink related threads.
>
> Not really, handlers are generic and created in ofproto/ofproto-dpif-upcall.c.
> We just don't have a recv() callback for them to use in dpif-netdev, AFAICT.

ACK, looks like my memory is not in a good state :) Thanks for resetting it...

_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to