Re: udp: Question about busy_poll change

[Edward Cree <ecree@solarflare.com>]

On 09/04/14 15:51, Shawn Bohrer wrote:
> I believe the sfc case where you only have a single NAPI context is
> also valid and it seems reasonable to me that if you can detect that
> specific case that busy polling could be allowed.  I'm not sure how to
> detect this.  I'm sure patches are welcome.
I think that to detect this we would have to have (a) a flag in the
struct sock to say that "this socket is bound to a real address" (ie.
not INADDR_ANY, multicast or anycast), (b) a flag in the skb to say that
"the driver that received this packet only uses one NAPI context per
intf.  Then if a && b you can fill in sk_napi_id even on unconnected
sockets.  If this sounds reasonable, I'll put a patch together.
> If we are spinning on a NAPI context and a packet arrives in a
> different rx queue then you'll get unpredictable latencies and
> out of order packets.  For the people using this feature that is
> probably not desirable.
A further question is, what happens if the user removes the IP address
from one interface and adds it to another?  AFAICT they will keep
busy_polling the old NAPI context until a packet is received on the new
interface and updates sk->sk_napi_id; and this is still the case even if
the socket is connected.  Of course this could be a case of "don't do
that then", but I can imagine some kind of hot-failover setup doing this.
This is also why we'd need a flag in the skb and can't just find out at
bind() time what the driver is and whether it uses multiple NAPI
contexts - because when the socket moves interface, it could end up on a
different driver.