Re: [net-next,v3] tcp: Improve setsockopt() TCP_USER_TIMEOUT accuracy

On 07/10/2018 05:33 PM, Jonathan Maxwell wrote:

On Tue, Jul 10, 2018 at 10:48 PM, Eric Dumazet [off-list ref] wrote:

quoted

On 07/10/2018 05:38 AM, Eric Dumazet wrote:

quoted

Note that if we always do jiffies_to_msecs(icsk->icsk_user_timeout) in TCP,
we also could change the convention and store msecs in this field instead of jiffies.

That would eliminate the msecs_to_jiffies() and jiffies_to_msecs() dance.

(That would be done in a patch of its own, of course)

tcp_keepalive_timer() does use icsk->icsk_user_timeout directly in jiffies unit,
but considering keeapalive timers are rarely used, this point would have to
do the msecs_to_jiffies() conversion.

and also if icsk->icsk_user_timeout = 0, then timeout in retransmits_timed_out()
is in jiffies and that would need to addressed.

Absolutely, this is what I was suggesting.

Pseudo code for this part, before your changes.

diff --git a/net/ipv4/tcp_timer.c b/net/ipv4/tcp_timer.c
index 3b3611729928f77934e0298bb248e55c7a7c5def..cae7bbc956ed51e9d381650957f54550cc0967d9 100644
--- a/net/ipv4/tcp_timer.c
+++ b/net/ipv4/tcp_timer.c

@@ -183,8 +183,9 @@ static bool retransmits_timed_out(struct sock *sk,
                else
                        timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
                                (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
+               timeout = jiffies_to_msecs(timeout)'
        }
-       return (tcp_time_stamp(tcp_sk(sk)) - start_ts) >= jiffies_to_msecs(timeout);
+       return (tcp_time_stamp(tcp_sk(sk)) - start_ts) >= timeout;
 }
 
 /* A write timeout has occurred. Process the after effects. */