Re: [PATCH net v3] net/sched: cbs: perform rate conversions in signed 64-bit arithmetic

[Paolo Abeni <pabeni@redhat.com>]

On 8/13/26 8:02 AM, Jack Wang wrote:
> cbs_set_port_rate() and cbs_change() multiply link rates and slope values
> by BYTES_PER_KBIT, an unsigned long constant.  On 32-bit architectures,
> the multiplications therefore take place in 32-bit unsigned arithmetic
> before the results are assigned to s64 fields.
> 
> For port rates above approximately 34.36 Gbit/s this wraps port_rate.
> The same conversion turns a negative sendslope into a large positive
> value, reversing the CBS credit adjustment.  This affects software CBS;
> port_rate is also refreshed on NETDEV_UP and NETDEV_CHANGE notifications.
> 
> Cast the first operand of each multiplication to s64 so all intermediate
> operations use signed 64-bit arithmetic and preserve the value's sign on
> every architecture
> 
> Also reject a negative idleslope. A negative idleslope can arm the
> watchdog in the past and busy-loop.
> 
> Fixes: 585d763af09c ("net/sched: Introduce Credit Based Shaper (CBS) qdisc")
> Fixes: 397006ba5d918 ("net/sched: cbs: Fix integer overflow in cbs_set_port_rate()")
> 
> Signed-off-by: Jack Wang <163wangjack@gmail.com>

No empty lines in the tag area.

> ---
> v3:
> - Reject only negative idleslope values.
> - Keep idleslope 0 accepted for compatibility with existing tc-testing
>   defaults, fixing test 1820 regression reported by Victor.
> - Keep the timediff_to_credits() overflow out of this series as a separate
>   follow-up.
> 
> v2:
> - Reject non-positive idleslope values to prevent scheduling the watchdog
>   in the past.
> 
> v1: https://lore.kernel.org/netdev/20260806155253.50252-1-163wangjack@gmail.com/
> 
>  net/sched/sch_cbs.c | 10 +++++++---
>  1 file changed, 7 insertions(+), 3 deletions(-)
> 
> diff --git a/net/sched/sch_cbs.c b/net/sched/sch_cbs.c
> index 1c93469c56e3..8db98d7c98a8 100644
> --- a/net/sched/sch_cbs.c
> +++ b/net/sched/sch_cbs.c
> @@ -335,7 +335,7 @@ static void cbs_set_port_rate(struct net_device *dev, struct cbs_sched_data *q)
>  		speed = ecmd.base.speed;
>  
>  skip:
> -	port_rate = speed * 1000 * BYTES_PER_KBIT;
> +	port_rate = (s64)speed * 1000 * BYTES_PER_KBIT;
>  
>  	atomic64_set(&q->port_rate, port_rate);
>  	netdev_dbg(dev, "cbs: set %s's port_rate to: %lld, linkspeed: %d\n",
> @@ -392,6 +392,10 @@ static int cbs_change(struct Qdisc *sch, struct nlattr *opt,
>  	}
>  
>  	qopt = nla_data(tb[TCA_CBS_PARMS]);
> +	if (qopt->idleslope < 0) {
> +		NL_SET_ERR_MSG(extack, "Idleslope must not be negative");

Sashiko noted this isn't enough to avoid the busy loop timer:

https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260813060256.28748-1-163wangjack%40gmail.com

I understand that a 0 value must be accepted, but I think we need either
a better solution or drop entirely this check that does not provide much
value per se. I think the first option would be better.

Also please note saskiko has more comments.

/P