On Tue, 8 Dec 2020 17:19:10 +0800 Cambda Zhu wrote:
> For each TCP zero window probe, the icsk_backoff is increased by one and
> its max value is tcp_retries2. If tcp_retries2 is greater than 63, the
> probe0 timeout shift may exceed its max bits. On x86_64/ARMv8/MIPS, the
> shift count would be masked to range 0 to 63. And on ARMv7 the result is
> zero. If the shift count is masked, only several probes will be sent
> with timeout shorter than TCP_RTO_MAX. But if the timeout is zero, it
> needs tcp_retries2 times probes to end this false timeout. Besides,
> bitwise shift greater than or equal to the width is an undefined
> behavior.
If icsk_backoff can reach 64, can it not also reach 256 and wrap?
Adding Eric's address from MAINTAINERS to CC.
> This patch adds a limit to the backoff. The max value of max_when is
> TCP_RTO_MAX and the min value of timeout base is TCP_RTO_MIN. The limit
> is the backoff from TCP_RTO_MIN to TCP_RTO_MAX.
> diff --git a/include/net/tcp.h b/include/net/tcp.h
> index d4ef5bf94168..82044179c345 100644
> --- a/include/net/tcp.h
> +++ b/include/net/tcp.h
> @@ -1321,7 +1321,9 @@ static inline unsigned long tcp_probe0_base(const
> struct sock *sk)
> static inline unsigned long tcp_probe0_when(const struct sock *sk,
> unsigned long max_when)
> {
> - u64 when = (u64)tcp_probe0_base(sk) << inet_csk(sk)->icsk_backoff;
> + u8 backoff = min_t(u8, ilog2(TCP_RTO_MAX / TCP_RTO_MIN) + 1,
> + inet_csk(sk)->icsk_backoff);
> + u64 when = (u64)tcp_probe0_base(sk) << backoff;
>
> return (unsigned long)min_t(u64, when, max_when);
> }
