Exiting from loss recovery

Randall Stewart rrs at netflix.com
Mon Oct 10 13:56:28 UTC 2016


I have a bit of experience now with this code.

If you exit recovery where Lawerence has marked, then when you go
down a few lines and credit the ack to the cwnd it can be a very
large “stretch” ack.. making cwnd spike up incorrectly by standard
new-reno terms… 

If you do move exit recover to there, you probably need to have a limit
on how big the cwnd can move… i.e. some sort of segment limit.


> On Oct 6, 2016, at 8:22 PM, hiren panchasara <hiren at strugglingcoder.info> wrote:
> In tcp_do_segment():
>                /*
>                 * If the congestion window was inflated to account
>                 * for the other side's cached packets, retract it.
>                 */
>                if (IN_FASTRECOVERY(tp->t_flags)) {
>                        if (SEQ_LT(th->th_ack, tp->snd_recover)) {
>                                if (tp->t_flags & TF_SACK_PERMIT)
>                                        tcp_sack_partialack(tp, th); 
>                                else
>                                        tcp_newreno_partial_ack(tp, th); 
>                        } else 
>                                cc_post_recovery(tp, th); 
>                }
> Here, if we get an ack that marks recovery from loss i.e. >=
> snd_recovery, we call cc_post_recovery() which in-turn calls CC specific
> post_recovery routine. But we don't reset TF_FASTRECOVERY |
> Later in the code we do this check again in 'process_ACK:'
>                /* XXXLAS: Can this be moved up into cc_post_recovery? */
>                if (IN_RECOVERY(tp->t_flags) &&
>                    SEQ_GEQ(th->th_ack, tp->snd_recover)) {
>                        EXIT_RECOVERY(tp->t_flags);
>                }
> And as it can be seen, Lawrence marked it as something that could
> possibly be done here and at the end of cc_post_recovery(). 
> So, should we do it? i.e call EXIT_RECOVERY() at the end of
> cc_post_recovery() and remove the block from 'process_ACK' section? or
> there is something subtle I am not seeing?
> Cheers,
> Hiren

Randall Stewart
rrs at netflix.com

More information about the freebsd-transport mailing list