On Mon, 2015-02-16 at 13:38 +0800, Fan Du wrote:
于 2015年02月13日 20:31, Eric Dumazet 写道:
quoted
We certainly wont add yet another timer in tcp socket for such usage.
And a buggy one, since you forgot all the refcounting associated with
such timers.
oh, embarrassing...
Will place probe timer aside with icsk_delack_timer in struct inet_connection_sock,
and manipulate through sk_reset_timer.
No : I said we would _not_ accept yet another timer (ie a big structure)
for such very rare event.
You can implement a pseudo timer, using a simple u32 field, that you
test in tcp_mtu_probe() to clear icsk->icsk_mtup.probe_size when enough
time has elapsed.
(tcp_time_stamp is u32, so are tp->lsndtime, tp->snd_cwnd_stamp,
tp->rcv_tstamp, tp->retrans_stamp, tp->tso_deferred, ...)