On Wed, May 07, 2025 at 07:26:03PM +0800, Zqiang wrote:
> For built with CONFIG_PREEMPT_RT=y kernels, running rcutorture
> tests resulted in the following splat:
>
> [ 68.797425] rcutorture_one_extend_check during change: Current 0x1 To add
> 0x1 To remove 0x0 preempt_count() 0x0
> [ 68.797533] WARNING: CPU: 2 PID: 512 at kernel/rcu/rcutorture.c:1993
> rcutorture_one_extend_check+0x419/0x560 [rcutorture]
> [ 68.797601] Call Trace:
> [ 68.797602] <TASK>
> [ 68.797619] ? lockdep_softirqs_off+0xa5/0x160
> [ 68.797631] rcutorture_one_extend+0x18e/0xcc0 [rcutorture
> 2466dbd2ff34dbaa36049cb323a80c3306ac997c]
> [ 68.797646] ? local_clock+0x19/0x40
> [ 68.797659] rcu_torture_one_read+0xf0/0x280 [rcutorture
> 2466dbd2ff34dbaa36049cb323a80c3306ac997c]
> [ 68.797678] ? __pfx_rcu_torture_one_read+0x10/0x10 [rcutorture
> 2466dbd2ff34dbaa36049cb323a80c3306ac997c]
> [ 68.797804] ? __pfx_rcu_torture_timer+0x10/0x10 [rcutorture
> 2466dbd2ff34dbaa36049cb323a80c3306ac997c]
> [ 68.797815] rcu-torture: rcu_torture_reader task started
> [ 68.797824] rcu-torture: Creating rcu_torture_reader task
> [ 68.797824] rcu_torture_reader+0x238/0x580 [rcutorture
> 2466dbd2ff34dbaa36049cb323a80c3306ac997c]
> [ 68.797836] ? kvm_sched_clock_read+0x15/0x30
>
> Disable BH does not change the SOFTIRQ corresponding bits in
> preempt_count() for RT kernels, this commit therefore use
> softirq_count() to check the if BH is disabled.
>
> Signed-off-by: Zqiang <qiang.zhang1...@gmail.com>
> ---
> kernel/rcu/rcutorture.c | 9 ++++++---
> 1 file changed, 6 insertions(+), 3 deletions(-)
>
> diff --git a/kernel/rcu/rcutorture.c b/kernel/rcu/rcutorture.c
> index 373c65a6e103..ef439569f979 100644
> --- a/kernel/rcu/rcutorture.c
> +++ b/kernel/rcu/rcutorture.c
> @@ -471,7 +471,7 @@ rcu_read_delay(struct torture_random_state *rrsp, struct
> rt_read_seg *rtrsp)
> !(torture_random(rrsp) % (nrealreaders * 2000 * longdelay_ms))) {
> started = cur_ops->get_gp_seq();
> ts = rcu_trace_clock_local();
> - if (preempt_count() & (SOFTIRQ_MASK | HARDIRQ_MASK))
> + if ((preempt_count() & HARDIRQ_MASK) || softirq_count())
> longdelay_ms = 5; /* Avoid triggering BH limits. */
> mdelay(longdelay_ms);
> rtrsp->rt_delay_ms = longdelay_ms;
> @@ -1990,7 +1990,7 @@ static void rcutorture_one_extend_check(char *s, int
> curstate, int new, int old,
> return;
>
> WARN_ONCE((curstate & (RCUTORTURE_RDR_BH | RCUTORTURE_RDR_RBH)) &&
> - !(preempt_count() & SOFTIRQ_MASK), ROEC_ARGS);
> + !softirq_count(), ROEC_ARGS);
> WARN_ONCE((curstate & (RCUTORTURE_RDR_PREEMPT | RCUTORTURE_RDR_SCHED))
> &&
> !(preempt_count() & PREEMPT_MASK), ROEC_ARGS);
> WARN_ONCE(cur_ops->readlock_nesting &&
> @@ -2004,7 +2004,7 @@ static void rcutorture_one_extend_check(char *s, int
> curstate, int new, int old,
>
> WARN_ONCE(cur_ops->extendables &&
> !(curstate & (RCUTORTURE_RDR_BH | RCUTORTURE_RDR_RBH)) &&
> - (preempt_count() & SOFTIRQ_MASK), ROEC_ARGS);
> + softirq_count(), ROEC_ARGS);
Given that softirq_count is defined as (preempt_count() & SOFTIRQ_MASK)
for CONFIG_PREEMPT_RT=n, the above don't change anything in that case,
so good. For CONFIG_PREEMPT_RT=y, softirq_count() looks to be the way
to check BH-disable nesting, so that is good as well.
> /*
> * non-preemptible RCU in a preemptible kernel uses preempt_disable()
> @@ -2025,6 +2025,9 @@ static void rcutorture_one_extend_check(char *s, int
> curstate, int new, int old,
> if (!IS_ENABLED(CONFIG_PREEMPT_RCU))
> mask |= RCUTORTURE_RDR_PREEMPT | RCUTORTURE_RDR_SCHED;
>
> + if (IS_ENABLED(CONFIG_PREEMPT_RT) && softirq_count())
> + mask |= RCUTORTURE_RDR_BH | RCUTORTURE_RDR_RBH;
At this point in the code, we are complaining if something is disabled
when it is not supposed to be. So if I understand this correctly, this
added code would suppress complaints (but only in CONFIG_PREEMPT_RT=y
kernels) when there is an unexpected rcu_read_lock() in the case where
there was either local_bh_disable() or rcu_read_lock_bh() in effect.
So I would expect that the CONFIG_PREEMPT_RT=y version of both
local_bh_disable() and rcu_read_lock_bh() would contain rcu_read_lock().
And in fact, rcu_read_lock_bh() invokes local_bh_disable(),
which, for CONFIG_PREEMPT_RT=y invokes __local_bh_disable_ip() in
kernel/softirq.c, which on the outermost local_bh_disabe() really does
invoke rcu_read_lock().
So this one looks good as well!
Reviewed-by: Paul E. McKenney <paul...@kernel.org>
> +
> WARN_ONCE(cur_ops->readlock_nesting && !(curstate & mask) &&
> cur_ops->readlock_nesting() > 0, ROEC_ARGS);
> }
> --
> 2.17.1
>
