Re: [PATCH net] tcp: reset tp->sacked_out when sack is enabled

[Eric Dumazet]

On Wed, Oct 26, 2022 at 7:12 AM Lu Wei <luwei32@xxxxxxxxxx> wrote:
>
> The meaning of tp->sacked_out depends on whether sack is enabled
> or not. If setsockopt is called to enable sack_ok via
> tcp_repair_options_est(), tp->sacked_out should be cleared, or it
> will trigger warning in tcp_verify_left_out as follows:
>
> ============================================
> WARNING: CPU: 8 PID: 0 at net/ipv4/tcp_input.c:2132
> tcp_timeout_mark_lost+0x154/0x160
> tcp_enter_loss+0x2b/0x290
> tcp_retransmit_timer+0x50b/0x640
> tcp_write_timer_handler+0x1c8/0x340
> tcp_write_timer+0xe5/0x140
> call_timer_fn+0x3a/0x1b0
> __run_timers.part.0+0x1bf/0x2d0
> run_timer_softirq+0x43/0xb0
> __do_softirq+0xfd/0x373
> __irq_exit_rcu+0xf6/0x140
>
> This warning occurs in several steps:
> Step1. If sack is not enabled, when server receives dup-ack,
>        it calls tcp_add_reno_sack() to increase tp->sacked_out.
>
> Step2. Setsockopt() is called to enable sack
>
> Step3. The retransmit timer expires, it calls tcp_timeout_mark_lost()
>        to increase tp->lost_out but not clear tp->sacked_out because
>        sack is enabled and tcp_is_reno() is false.
>
> So tp->left_out is increased repeatly in Step1 and Step3 and it is
> greater than tp->packets_out and trigger the warning. In function
> tcp_timeout_mark_lost(), tp->sacked_out will be cleared if Step2 not
> happen and the warning will not be triggered. So this patch clears
> tp->sacked_out in tcp_repair_options_est().
>
> Fixes: b139ba4e90dc ("tcp: Repair connection-time negotiated parameters")
> Signed-off-by: Lu Wei <luwei32@xxxxxxxxxx>
> ---
>  net/ipv4/tcp.c | 3 +++
>  1 file changed, 3 insertions(+)
>
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index ef14efa1fb70..188d5c0e440f 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -3282,6 +3282,9 @@ static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
>                         if (opt.opt_val != 0)
>                                 return -EINVAL;
>
> +                       if (tcp_is_reno(tp))
> +                               tp->sacked_out = 0;
> +
>                         tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
>                         break;
>                 case TCPOPT_TIMESTAMP:
> --
> 2.31.1
>

Hmm, I am not sure this is the right fix.

Probably TCP_REPAIR_OPTIONS should not be allowed if data has already been sent.

Pavel, what do you think ?