Le Sun, Oct 20, 2024 at 08:51:19PM +0800, Zqiang a écrit :
> Currently, running rcutorture test with torture_type=rcu fwd_progress=8
> n_barrier_cbs=8 nocbs_nthreads=8 nocbs_toggle=100 onoff_interval=60
> test_boost=2, will trigger the following warning:
>
> WARNING: CPU: 19 PID: 100 at kernel/rcu/tree_nocb.h:1061 rcu_nocb_rdp_deoffload+0x292/0x2a0
> RIP: 0010:rcu_nocb_rdp_deoffload+0x292/0x2a0
> [18839.537322] Call Trace:
> [18839.538006] <TASK>
> [18839.538596] ? __warn+0x7e/0x120
> [18839.539491] ? rcu_nocb_rdp_deoffload+0x292/0x2a0
> [18839.540757] ? report_bug+0x18e/0x1a0
> [18839.541805] ? handle_bug+0x3d/0x70
> [18839.542837] ? exc_invalid_op+0x18/0x70
> [18839.543959] ? asm_exc_invalid_op+0x1a/0x20
> [18839.545165] ? rcu_nocb_rdp_deoffload+0x292/0x2a0
> [18839.546547] rcu_nocb_cpu_deoffload+0x70/0xa0
> [18839.547814] rcu_nocb_toggle+0x136/0x1c0
> [18839.548960] ? __pfx_rcu_nocb_toggle+0x10/0x10
> [18839.550073] kthread+0xd1/0x100
> [18839.550958] ? __pfx_kthread+0x10/0x10
> [18839.552008] ret_from_fork+0x2f/0x50
> [18839.553002] ? __pfx_kthread+0x10/0x10
> [18839.553968] ret_from_fork_asm+0x1a/0x30
> [18839.555038] </TASK>
>
> CPU0 CPU2 CPU3
> //rcu_nocb_toggle //nocb_cb_wait //rcutorture
>
> // deoffload CPU1 // process CPU1's rdp
> rcu_barrier()
> rcu_segcblist_entrain()
> rcu_segcblist_add_len(1);
> // len == 2
> // enqueue barrier
> // callback to CPU1's
> // rdp->cblist
> rcu_do_batch()
> // invoke CPU1's rdp->cblist
> // callback
> rcu_barrier_callback()
> rcu_barrier()
> mutex_lock(&rcu_state.barrier_mutex);
> // still see len == 2
> // enqueue barrier callback
> // to CPU1's rdp->cblist
> rcu_segcblist_entrain()
> rcu_segcblist_add_len(1);
> // len == 3
> // decrement len
> rcu_segcblist_add_len(-2);
> kthread_parkme()
>
> // CPU1's rdp->cblist len == 1
> // Warn because there is
> // still a pending barrier
> // trigger warning
> WARN_ON_ONCE(rcu_segcblist_n_cbs(&rdp->cblist));
> cpus_read_unlock();
>
> // wait CPU1 comes online
> // invoke barrier callback on
> // CPU1 rdp's->cblist
> wait_for_completion(&rcu_state.barrier_completion);
> // deoffload CPU4
> cpus_read_lock()
> rcu_barrier()
> mutex_lock(&rcu_state.barrier_mutex);
> // block on barrier_mutex
> // wait rcu_barrier() on
> // CPU3 to unlock barrier_mutex
> // but CPU3 unlock barrier_mutex
> // need to wait CPU1 comes online
> // when CPU1 going online will block on cpus_write_lock
>
> The above scenario will not only trigger WARN_ON_ONCE(), but also
> trigger deadlock, this commit therefore check rdp->cblist length
> before invoke kthread_parkme(), and the kthread_parkme() is not
> invoke until length reaches zero.
>
> Signed-off-by: Zqiang <qiang.zhang1211@xxxxxxxxx>
> ---
> kernel/rcu/tree_nocb.h | 8 +++++++-
> 1 file changed, 7 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
> index 8648233e1717..a2b0ebdefee3 100644
> --- a/kernel/rcu/tree_nocb.h
> +++ b/kernel/rcu/tree_nocb.h
> @@ -893,6 +893,12 @@ static inline bool nocb_cb_wait_cond(struct rcu_data *rdp)
> return !READ_ONCE(rdp->nocb_cb_sleep) || kthread_should_park();
> }
>
> +static inline bool nocb_cblist_empty(struct rcu_data *rdp)
> +{
> + return !(rcu_rdp_is_offloaded(rdp) &&
But the rdp has to be offloaded when nocb_cb_wait() is running, and that
include the times when it is parking and when it is unparking.
> + WARN_ON_ONCE(rcu_segcblist_n_cbs(&rdp->cblist)));
And like your scenario above shows, it's possible to reach here with
callbacks. So this check shouldn't be a warning at that point?
> +}
> +
> /*
> * Invoke any ready callbacks from the corresponding no-CBs CPU,
> * then, if there are no more, wait for more to appear.
> @@ -907,7 +913,7 @@ static void nocb_cb_wait(struct rcu_data *rdp)
>
> swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
> nocb_cb_wait_cond(rdp));
> - if (kthread_should_park()) {
> + if (kthread_should_park() && nocb_cblist_empty(rdp)) {
What about this instead? If the second barrier is queued before
the final check to rcu_segcblist_ready_cbs() in nocb_cb_wait(), this
will be noticed and ->nocb_cb_sleep will remain false. If otherwise rcu_barrier()
is called after that final rcu_segcblist_ready_cbs() check, it will observe
the final decrement to zero and won't entrain the callback.
diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
index 16865475120b..0de07d44646c 100644
--- a/kernel/rcu/tree_nocb.h
+++ b/kernel/rcu/tree_nocb.h
@@ -891,7 +891,19 @@ static void nocb_cb_wait(struct rcu_data *rdp)
swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
nocb_cb_wait_cond(rdp));
if (kthread_should_park()) {
- kthread_parkme();
+ /*
+ * kthread_park() must be preceded by an rcu_barrier().
+ * But yet another rcu_barrier() might have sneaked in between
+ * the barrier callback execution and the callbacks counter
+ * decrement.
+ */
+ if (rdp->nocb_cb_sleep) {
+ rcu_nocb_lock_irqsave(rdp, flags);
+ WARN_ON_ONCE(rcu_segcblist_n_cbs(&rdp->cblist));
+ rcu_nocb_unlock_irqrestore(rdp, flags);
+
+ kthread_parkme();
+ }
} else if (READ_ONCE(rdp->nocb_cb_sleep)) {
WARN_ON(signal_pending(current));
trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
