On Wed, 2016-03-30 at 21:05 +0000, Light, John J wrote:

David,

I see a recent change in inet_connection_sock.c that uses
sysctl_tcp_synack_retries suspiciously.

Previously, sysctl_tcp_synack_retries was a global variable, and now
it has been moved into the fragment (really netns_ipv4).

You mean the network namespace ?

In reqsk_timer_handler (inet_connection_sock.c) near line 563 it is
used as the default max_retries value in case icsk->icsk_syn_retries
is not set.

Previously, other transports (dccp?) would use the TCP global variable
as a default.  Now that the defaults come from the fragment, I suspect
the fragment variable is not set for other transports.  (I can't find
where it's set for dccp.)

Of course, this will only show in non-TCP transport protocols when the
icsk retries value is not set, so it's a rare case.  Perhaps it's an
unreachable case, since I don't know all the kernel paths.

Maybe the problem is that the default shouldn't be to a TCP value, but
should be a 'transport' value.

This code is somewhat convoluted, so I am not sure of my analysis, but
I wanted you to consider it.

No idea why you ask David instead of patch author ?

I don't see any problem here.

Do we really want to spend time on this minor issue ?