Hi,
I noticed that in tcp_mark_lost_retrans the for-loop is only entered
if tcp_is_fack(tp) evaluates to true:
if (!tcp_is_fack(tp) || !tp->retrans_out ||
!after(received_upto, tp->lost_retrans_low) ||
icsk->icsk_ca_state != TCP_CA_Recovery)
return;
Therefore the following check in the for-loop seems to be redundant,
because it always evaluates to true:
(tcp_is_fack(tp) ||
!before(received_upto,
ack_seq + tp->reordering * tp->mss_cache))
Did I miss something?
Best regards,
Arnd Hannemann
From: Arnd Hannemann <[email protected]>
Because the for loop is only executed for FACK-enabled flows remove
redundant checks within the loop.
Signed-off-by: Arnd Hannemann <[email protected]>
---
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 99b7ecb..cd8b4bd 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -1178,10 +1178,7 @@ static void tcp_mark_lost_retrans(struct sock *sk)
if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
continue;
- if (after(received_upto, ack_seq) &&
- (tcp_is_fack(tp) ||
- !before(received_upto,
- ack_seq + tp->reordering * tp->mss_cache))) {
+ if (after(received_upto, ack_seq)) {
TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
tp->retrans_out -= tcp_skb_pcount(skb);
On Wed, 7 Jan 2009, Arnd Hannemann wrote:
> I noticed
Good that somebody else is looking TCP code besides me... :-)
> that in tcp_mark_lost_retrans the for-loop is only entered
> if tcp_is_fack(tp) evaluates to true:
>
> if (!tcp_is_fack(tp) || !tp->retrans_out ||
> !after(received_upto, tp->lost_retrans_low) ||
> icsk->icsk_ca_state != TCP_CA_Recovery)
> return;
>
> Therefore the following check in the for-loop seems to be redundant,
> because it always evaluates to true:
>
> (tcp_is_fack(tp) ||
> !before(received_upto,
> ack_seq + tp->reordering * tp->mss_cache))
>
> Did I miss something?
It was just a left over from the RFC3517 SACK addition which added that
!tcp_is_fack(tp) there above. ...It would have been nice to have similar
lost rexmit feature without FACK as well but calculating that wasn't
trivial (or I didn't find that too trivial) and could end up being
extremely expensive in case of large holes. (So I also left it there as
sort of reminder).
On the second thought, it would be possible to count skbs we pass while
walking from the beginning and use that a remaining_sacked counter
to get rid of all heurestics too and base the counting only on sacked
stuff which aligns with the spirit of rfc3517 much better than
sacked+holes used by fack.
> Best regards,
> Arnd Hannemann
>
> From: Arnd Hannemann <[email protected]>
>
> Because the for loop is only executed for FACK-enabled flows remove
> redundant checks within the loop.
>
> Signed-off-by: Arnd Hannemann <[email protected]>
> ---
> diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> index 99b7ecb..cd8b4bd 100644
> --- a/net/ipv4/tcp_input.c
> +++ b/net/ipv4/tcp_input.c
> @@ -1178,10 +1178,7 @@ static void tcp_mark_lost_retrans(struct sock *sk)
> if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
> continue;
>
> - if (after(received_upto, ack_seq) &&
> - (tcp_is_fack(tp) ||
> - !before(received_upto,
> - ack_seq + tp->reordering * tp->mss_cache))) {
> + if (after(received_upto, ack_seq)) {
> TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
> tp->retrans_out -= tcp_skb_pcount(skb);
--
i.
On Wed, 7 Jan 2009, Ilpo J?rvinen wrote:
> On Wed, 7 Jan 2009, Arnd Hannemann wrote:
>
> > I noticed
>
> Good that somebody else is looking TCP code besides me... :-)
>
> > that in tcp_mark_lost_retrans the for-loop is only entered
> > if tcp_is_fack(tp) evaluates to true:
> >
> > if (!tcp_is_fack(tp) || !tp->retrans_out ||
> > !after(received_upto, tp->lost_retrans_low) ||
> > icsk->icsk_ca_state != TCP_CA_Recovery)
> > return;
> >
> > Therefore the following check in the for-loop seems to be redundant,
> > because it always evaluates to true:
> >
> > (tcp_is_fack(tp) ||
> > !before(received_upto,
> > ack_seq + tp->reordering * tp->mss_cache))
> >
> > Did I miss something?
>
> It was just a left over from the RFC3517 SACK addition which added that
> !tcp_is_fack(tp) there above. ...It would have been nice to have similar
> lost rexmit feature without FACK as well but calculating that wasn't
> trivial (or I didn't find that too trivial) and could end up being
> extremely expensive in case of large holes. (So I also left it there as
> sort of reminder).
>
> On the second thought, it would be possible to count skbs we pass while
> walking from the beginning and use that a remaining_sacked counter
> to get rid of all heurestics too and base the counting only on sacked
> stuff which aligns with the spirit of rfc3517 much better than
> sacked+holes used by fack.
Nah, tried to do that that wasn't working nicely either... Since there is
a need to know how many sack blocks reside between ack_seq and
received_upto, not the number of sack blocks between skb and
received_upto...
--
i.
Ilpo J?rvinen schrieb:
> On Wed, 7 Jan 2009, Arnd Hannemann wrote:
>
>
>> I noticed
>>
>
> Good that somebody else is looking TCP code besides me... :-)
>
Well I try hard... ;-)
>
>> that in tcp_mark_lost_retrans the for-loop is only entered
>> if tcp_is_fack(tp) evaluates to true:
>>
>> if (!tcp_is_fack(tp) || !tp->retrans_out ||
>> !after(received_upto, tp->lost_retrans_low) ||
>> icsk->icsk_ca_state != TCP_CA_Recovery)
>> return;
>>
>> Therefore the following check in the for-loop seems to be redundant,
>> because it always evaluates to true:
>>
>> (tcp_is_fack(tp) ||
>> !before(received_upto,
>> ack_seq + tp->reordering * tp->mss_cache))
>>
>> Did I miss something?
>>
>
> It was just a left over from the RFC3517 SACK addition which added that
> !tcp_is_fack(tp) there above. ...It would have been nice to have similar
> lost rexmit feature without FACK as well but calculating that wasn't
> trivial (or I didn't find that too trivial) and could end up being
> extremely expensive in case of large holes. (So I also left it there as
> sort of reminder).
>
Perhaps it would be better to let the comments reflect
what you just said and remove the redundant check
anyway to reduce the dead code a newcomer has to understand ;-)
I would have included a patch for the comments, but as you have a
deeper understanding of the code it would probably
be better if you can do it.
Best regards,
Arnd
On Wed, 7 Jan 2009, Arnd Hannemann wrote:
> Perhaps it would be better to let the comments reflect
> what you just said and remove the redundant check
> anyway to reduce the dead code a newcomer has to understand ;-)
> I would have included a patch for the comments, but as you have a
> deeper understanding of the code it would probably
> be better if you can do it.
I have a patch for this in a local tree and will submit it among
other TCP stuff (later on).
--
i.