Re: [PATCH v2 net] tcp: zero retrans_stamp if all retrans were acked

[Marcelo Ricardo Leitner <mleitner@redhat.com>]

On 04-11-2014 23:20, Eric Dumazet wrote:
> On Tue, 2014-11-04 at 18:51 -0200, Marcelo Ricardo Leitner wrote:
>
>> And thank you guys for all the assistance on it. Btw, would you send me that
>> packetdrill script? I'm curious to see how such corner case could be written
>> on it.
>
> One of the script I saw was :
>
> You might have to adapt preconditions (tcp_rmem[]/tcp_wmem[])
>
> 0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
> // Set a 10s timeout
> +.000 setsockopt(3, SOL_TCP, TCP_USER_TIMEOUT, [10000], 4) = 0
> +.000 bind(3, ..., ...) = 0
> +.000 listen(3, 1) = 0
> +.000 < S 0:0(0) win 32792 <mss 1460,nop,wscale 7>
> +.000 > S. 0:0(0) ack 1 <mss 1460,nop,wscale 6>
> +.010 < . 1:1(0) ack 1 win 257
> +.000 accept(3, ..., ...) = 4
> +.000 write(4, ..., 1000) = 1000
> +.000 > P. 1:1001(1000) ack 1
> +.625 > P. 1:1001(1000) ack 1
> +.020 < . 1:1(0) ack 1001 win 257
> // Purposely write more after the specified timeout for testing
> +11.0 write(4, ..., 1000) = 1000
> +.000 > P. 1001:2001(1000) ack 1
> +1.25 > P. 1001:2001(1000) ack 1
> // socket is killed when the 2nd RTO fires at +2.50 w/o this patch
> // so the next write returns ETIMEOUT
> +2.60 write(4, ..., 1000) = 1000

Cool, thanks!

Marcelo