sdf@xxxxxxxxxx <sdf@xxxxxxxxxx> [Tue, 2020-05-05 09:02 -0700]:
> On 05/04, Andrey Ignatov wrote:
> > Stanislav Fomichev <sdf@xxxxxxxxxx> [Mon, 2020-05-04 10:34 -0700]:
> > > We want to have a tighter control on what ports we bind to in
> > > the BPF_CGROUP_INET{4,6}_CONNECT hooks even if it means
> > > connect() becomes slightly more expensive. The expensive part
> > > comes from the fact that we now need to call inet_csk_get_port()
> > > that verifies that the port is not used and allocates an entry
> > > in the hash table for it.
>
> > FWIW: Initially that IP_BIND_ADDRESS_NO_PORT limitation came from the
> > fact that on my specific use-case (mysql client making 200-500 connects
> > per sec to mysql server) disabling the option was making application
> > pretty much unusable (inet_csk_get_port was taking more time than mysql
> > client connect timeout == 3sec).
>
> > But I guess for some use-cases that call sys_connect not too often it
> > makes sense.
> Yeah, I don't think we plan to reach those QPS numbers.
> But, for the record, did you try to bind to a random port in that
> case? And did you bail out on error or did a couple of retries?
Random port.
As for retries: no retries on low-level (no reconnecting to that same
server if sys_connect failed), but I don't quite remember how
higher-level behaved (it was choosing a server to connect to according
to some sharding scheme and I don't remember whether it was trying to
connect to next replica or not if current replica failed), sorry.
--
Andrey Ignatov
