On Tue, Sep 17, 2019 at 1:22 PM Eric Dumazet <eduma...@google.com> wrote:
>
> Tue, Sep 17, 2019 at 10:13 AM Jason Baron <jba...@akamai.com> wrote:
> >
> >
> > Hi,
> >
> > I was interested in adding a field to tcp_info around the TFO state of a
> > socket. So for the server side it would indicate if TFO was used to
> > create the socket and on the client side it would report whether TFO
> > worked and if not that it failed with maybe some additional states
> > around why it failed. I'm thinking it would be maybe 3 bits.
BTW, one aspect of that "did TFO work" info is available already in
tcp_info in the tcpi_options field.
Kernel side is:
if (tp->syn_data_acked)
info->tcpi_options |= TCPI_OPT_SYN_DATA;
We use this bit in packetdrill tests on client and server side to
check that the TFO data-in-SYN succeeded:
+0 %{ assert (tcpi_options & TCPI_OPT_SYN_DATA) != 0, tcpi_options }%
These TFO bits were added much later than the other bits, so IMHO it
would be OK to add more bits somewhere unused in tcp_info to indicate
reasons for TFO failure. Especially if, as you suggest, "0" as a code
point could indicate that the code point is undefined, and all
meaningful code points were non-zero.
neal
> > My question is whether its reasonable to use the unused bits of
> > __u8 tcpi_delivery_rate_app_limited:1;. Or is this not good because
> > the size hasn't changed? What if I avoided using 0 for the new field to
> > avoid the possibility of not knowing if 0 because its the old kernel or
> > 0 because that's now its a TFO state? IE the new field could always be >
> > 0 for the new kernel.
> >
>
> I guess that storing the 'why it has failed' would need more bits.
>
> I suggest maybe using an event for this, instead of TCP_INFO ?
>
> As of using the bits, maybe the monitoring application does not really care
> if running on an old kernel where the bits would be zero.
>
> Commit eb8329e0a04db0061f714f033b4454326ba147f4 reserved a single
> bit and did not bother about making sure the monitoring would detect if this
> runs on an old kernel.
