Re: [PATCH RFC] net/sched: taprio: enforce minimum software scheduling interval

[Uladzislau Zhauniarovich <uladzislau.zhauniarovich@gmail.com>]

The patch correctly identifies the failure mechanism: with software
scheduling, advance_sched() is a self-rearming hrtimer, and an interval
short enough re-arms it with an already-expired deadline so it fires
back to back in hardirq context, starving the RCU grace-period kthread.
Gating the new minimum on !FULL_OFFLOAD_IS_ENABLED(q->flags) is right —
fully offloaded schedules are advanced by the NIC and must keep the
link-speed-derived minimum. The (s64) cast in the cycle_time check and
the Fixes: b5b73b26b3ca tag are also correct. Keep all of that.

The chosen floor of 1 microsecond does not fix the bug class, only this
exact reproducer. The reproducer's 129 ns cycle is rejected, but the
cost of one advance_sched() invocation is on the order of 10
microseconds on the syzbot debug configuration (KASAN, lockdep): each
fire takes current_entry_lock, recomputes per-traffic-class budgets and
raises the TX softirq. Any cycle between 1 and ~10 microseconds still
passes the new validation and still re-arms the timer into the past,
reproducing the identical livelock. A trivially modified reproducer (or
the fuzzer itself) will reopen this bug as a new instance. The floor
must exceed the worst-case cost of servicing the timer with a safety
margin, not merely exceed hardware interrupt overhead as the commit
message currently argues.

Also note why validation passes at all: virtual devices report inflated
link speeds — veth advertises SPEED_10000 and bonding sums the speeds of
its members — so length_to_duration(q, ETH_ZLEN) drops to tens of
nanoseconds on the reproducer's bond0-over-veth topology. The commit
message should state this, since it explains why the existing
b5b73b26b3ca check is insufficient on virtual topologies.

Required corrections:

Raise the floor to 100 microseconds and rename the constant to
TAPRIO_MIN_SW_INTERVAL_NS, defined as (100 * NSEC_PER_USEC). Add a
comment above the definition explaining that the value must exceed the
cost of one advance_sched() invocation (lock acquisition, budget
recomputation, TX softirq) with margin, so the timer always leaves the
CPU idle time to make progress. A software schedule with sub-100us
entries has no legitimate use: the timer overhead alone exceeds the
gate interval.

Do not duplicate the max_t() clamping logic in fill_sched_entry()
and parse_taprio_schedule(). Introduce one small helper next to
length_to_duration(), e.g.:

static int taprio_min_interval(struct taprio_sched *q)
{
int min = length_to_duration(q, ETH_ZLEN);

  if (!FULL_OFFLOAD_IS_ENABLED(q->flags))
      min = max_t(int, min, TAPRIO_MIN_SW_INTERVAL_NS);

  return min;
}

and call it from both validation sites. Remove the bare { } block that
the current version inserts into parse_taprio_schedule(); with the
helper, the cycle_time check stays a single expression:


if (new->cycle_time < (s64)new->num_entries * taprio_min_interval(q)) {
3. Keep the (s64) cast on the num_entries multiplication, the
!FULL_OFFLOAD_IS_ENABLED() gating, the existing NL_SET_ERR_MSG texts,
and the Fixes: b5b73b26b3ca ("taprio: Fix allowing too small
intervals") tag.

Rework the commit message: (a) replace the 1us "interrupt overhead"
justification with the timer-service-cost argument above; (b) explain
that virtual devices defeat the link-speed minimum (veth reports 10
Gb/s, bonding sums member speeds, giving a ~24-48 ns minimum on the
reproducer topology); (c) state explicitly that fully offloaded
schedules are unaffected.

Verification data for the 100us value: an A/B run of the
tc-testing taprio suite (tools/testing/selftests/tc-testing,
tc-tests/qdiscs/taprio.json) against this floor shows all existing
cases still pass — every valid software schedule in the suite uses
300us or larger entries — while the reproducer's configuration is
rejected at qdisc creation with -EINVAL. So the stricter floor does not
regress any exercised configuration.

On 27/06/2026 00:22, syzbot wrote:
> When configuring taprio with a very small schedule interval (e.g., 129 ns),
> the kernel validates the interval against the time it takes to transmit a
> minimum-sized Ethernet frame (60 bytes). On high-speed links like 10 Gbps,
> this minimum duration is extremely small (e.g., 48 ns). Since the requested
> interval is larger than this, the validation passes.
>
> However, when hardware offload is not used, taprio falls back to software
> scheduling and arms an hrtimer. The hrtimer is programmed to fire every 129
> ns. This is significantly shorter than the overhead of handling a hardware
> interrupt and running the hrtimer subsystem. As a result, the timer
> constantly falls behind, and the CPU is livelocked in hardirq context
> endlessly servicing the advance_sched() hrtimer. This starves the RCU
> grace-period kthreads, leading to an RCU stall panic:
>
> rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
> rcu: 	1-...!: (1 GPs behind) idle=858c/1/0x4000000000000000
> softirq=112663/112663 fqs=0
> rcu: 	(detected by 0, t=10502 jiffies, g=143345, q=674 ncpus=2)
> Sending NMI from CPU 0 to CPUs 1:
> NMI backtrace for cpu 1
> ...
> Call Trace:
>   <IRQ>
>   advance_sched+0x99a/0xc80 net/sched/sch_taprio.c:988
>   __run_hrtimer kernel/time/hrtimer.c:1930 [inline]
>   __hrtimer_run_queues+0x3bc/0xa10 kernel/time/hrtimer.c:1994
>   hrtimer_interrupt+0x448/0x910 kernel/time/hrtimer.c:2113
>   local_apic_timer_interrupt arch/x86/kernel/apic/apic.c:1050 [inline]
>   __sysvec_apic_timer_interrupt+0x102/0x430 arch/x86/kernel/apic/apic.c:1067
>   instr_sysvec_apic_timer_interrupt arch/x86/kernel/apic/apic.c:1061
>   [inline]
>   sysvec_apic_timer_interrupt+0xa1/0xc0 arch/x86/kernel/apic/apic.c:1061
>   </IRQ>
>
> To fix this, enforce a hard absolute minimum interval of 1 microsecond
> (NSEC_PER_USEC) for software-based scheduling, regardless of the link
> speed. Hardware-offloaded taprio can continue to support smaller intervals
> since it does not rely on the CPU's hrtimer. The same logic is applied to
> the overall cycle_time validation, casting num_entries to s64 to prevent
> potential integer overflow.
>
> Fixes: b5b73b26b3ca ("taprio: Fix allowing too small intervals")
> Assisted-by: Gemini:gemini-3.1-pro-preview syzbot
> Reported-by: syzbot+19d01f6082ec61dd45b2@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=19d01f6082ec61dd45b2
> Link: https://syzkaller.appspot.com/ai_job?id=e96ce5ef-50a8-4856-a518-279d5610b23c
> To: "David S. Miller" <davem@davemloft.net>
> To: "Eric Dumazet" <edumazet@google.com>
> To: "Jamal Hadi Salim" <jhs@mojatatu.com>
> To: "Jiri Pirko" <jiri@resnulli.us>
> To: "Jakub Kicinski" <kuba@kernel.org>
> To: <netdev@vger.kernel.org>
> To: "Paolo Abeni" <pabeni@redhat.com>
> To: "Vinicius Costa Gomes" <vinicius.gomes@intel.com>
> Cc: "Simon Horman" <horms@kernel.org>
> Cc: <linux-kernel@vger.kernel.org>
>
> ---
> diff --git a/net/sched/sch_taprio.c b/net/sched/sch_taprio.c
> index 45245157e..b5743a556 100644
> --- a/net/sched/sch_taprio.c
> +++ b/net/sched/sch_taprio.c
> @@ -49,6 +49,11 @@ static struct static_key_false taprio_have_working_mqprio;
>    */
>   #define TAPRIO_PICOS_PER_BYTE_MIN 17
>   
> +/* The software scheduler cannot sustain hrtimer intervals smaller than
> + * this without livelocking the CPU.
> + */
> +#define TAPRIO_MIN_SW_INTERVAL NSEC_PER_USEC
> +
>   struct sched_entry {
>   	/* Durations between this GCL entry and the GCL entry where the
>   	 * respective traffic class gate closes
> @@ -1041,6 +1046,9 @@ static int fill_sched_entry(struct taprio_sched *q, struct nlattr **tb,
>   	int min_duration = length_to_duration(q, ETH_ZLEN);
>   	u32 interval = 0;
>   
> +	if (!FULL_OFFLOAD_IS_ENABLED(q->flags))
> +		min_duration = max_t(int, min_duration, TAPRIO_MIN_SW_INTERVAL);
> +
>   	if (tb[TCA_TAPRIO_SCHED_ENTRY_CMD])
>   		entry->command = nla_get_u8(
>   			tb[TCA_TAPRIO_SCHED_ENTRY_CMD]);
> @@ -1054,7 +1062,8 @@ static int fill_sched_entry(struct taprio_sched *q, struct nlattr **tb,
>   			tb[TCA_TAPRIO_SCHED_ENTRY_INTERVAL]);
>   
>   	/* The interval should allow at least the minimum ethernet
> -	 * frame to go out.
> +	 * frame to go out. When software scheduling is used, the
> +	 * interval must also be at least the minimum hrtimer interval.
>   	 */
>   	if (interval < min_duration) {
>   		NL_SET_ERR_MSG(extack, "Invalid interval for schedule entry");
> @@ -1166,9 +1175,17 @@ static int parse_taprio_schedule(struct taprio_sched *q, struct nlattr **tb,
>   		new->cycle_time = cycle;
>   	}
>   
> -	if (new->cycle_time < new->num_entries * length_to_duration(q, ETH_ZLEN)) {
> -		NL_SET_ERR_MSG(extack, "'cycle_time' is too small");
> -		return -EINVAL;
> +	{
> +		int min_duration = length_to_duration(q, ETH_ZLEN);
> +
> +		if (!FULL_OFFLOAD_IS_ENABLED(q->flags))
> +			min_duration = max_t(int, min_duration,
> +					     TAPRIO_MIN_SW_INTERVAL);
> +
> +		if (new->cycle_time < (s64)new->num_entries * min_duration) {
> +			NL_SET_ERR_MSG(extack, "'cycle_time' is too small");
> +			return -EINVAL;
> +		}
>   	}
>   
>   	taprio_calculate_gate_durations(q, new);
>
>
> base-commit: 8cd9520d35a6c38db6567e97dd93b1f11f185dc6