[ovs-dev] [PATCH V2 01/10] netdev: Extend rx_recv to pass multiple packets.

[Pravin Shelar]

On Fri, Mar 21, 2014 at 2:48 PM, Thomas Graf <tgraf at redhat.com> wrote:
> On 03/21/2014 10:28 PM, Pravin Shelar wrote:
>
>> I think it is not possible. Why vswitchd would generate such action if
>> the packet is not going somewhere?
>
>
> [...]
>
>
>> right, there is memory leak for packet size less than eth-header.
>> unfortunately dp_netdev_port_input() would not know if packet is
>> deleted, since it can send it to odp-execute.
>> At this point there is three places where we delete packet:
>> 1. dp_netdev_port_input()
>> 2. odp action (drop)
>> 3. odp-actions callback.
>>
>> I do not think if we can centralize it.
>
>
> odp_execute_actions() and the callback can be made returning a bool
> indicating whether the buffer was consumed, much like the kernel
> rx handler.
>
> I think it's slightly fragile as-is. Definitely don't want to hold
> up the series because of this. I'm happy to propose this as a follow up
> as well.
>
OK, But passing around values and checking it is not efficient
specially for fast path loop.

>
>>>> -    retval = rx->netdev->netdev_class->rx_recv(rx, buffer);
>>>> +    retval = rx->netdev->netdev_class->rx_recv(rx, buffers, cnt);
>>>>        if (!retval) {
>>>>            COVERAGE_INC(netdev_received);
>>>
>>>
>>>
>>> What would be awesome here is a COVERAGE_ADD(netdev_batched, cnt).
>>
>>
>> I am not sure what do u mean, but we can get avg batch size by looking
>> at all packets recved and this count for given period of time.
>
>
> True, hit+miss/netdev_received will provide a similar average.