On Wed, Feb 24, 2021 at 10:37:09AM -0800, Paul E. McKenney wrote:
On Tue, Feb 23, 2021 at 01:09:59AM +0100, Frederic Weisbecker wrote:
Two situations can cause a missed nocb timer rearm:
- rdp(CPU A) queues its nocb timer. The grace period elapses before the timer get a chance to fire. The nocb_gp kthread is awaken by rdp(CPU B). The nocb_cb kthread for rdp(CPU A) is awaken and process the callbacks, again before the nocb_timer for CPU A get a chance to fire. rdp(CPU A) queues a callback and wakes up nocb_gp kthread, cancelling the pending nocb_timer without resetting the corresponding nocb_defer_wakeup.
As discussed offlist, expanding the above scenario results in this sequence of steps:
There are no callbacks queued for any CPU covered by CPU 0-2's ->nocb_gp_kthread.
CPU 0 enqueues its first callback with interrupts disabled, and thus must defer awakening its ->nocb_gp_kthread. It therefore queues its rcu_data structure's ->nocb_timer.
CPU 1, which shares the same ->nocb_gp_kthread, also enqueues a callback, but with interrupts enabled, allowing it to directly awaken the ->nocb_gp_kthread.
The newly awakened ->nocb_gp_kthread associates both CPU 0's and CPU 1's callbacks with a future grace period and arranges for that grace period to be started.
This ->nocb_gp_kthread goes to sleep waiting for the end of this future grace period.
This grace period elapses before the CPU 0's timer fires. This is normally improbably given that the timer is set for only one jiffy, but timers can be delayed. Besides, it is possible that kernel was built with CONFIG_RCU_STRICT_GRACE_PERIOD=y.
The grace period ends, so rcu_gp_kthread awakens the ->nocb_gp_kthread, which in turn awakens both CPU 0's and CPU 1's ->nocb_cb_kthread.
CPU 0's ->nocb_cb_kthread invokes its callback.
Note that neither kthread updated any ->nocb_timer state, so CPU 0's ->nocb_defer_wakeup is still set to either RCU_NOCB_WAKE or RCU_NOCB_WAKE_FORCE.
CPU 0 enqueues its second callback, again with interrupts disabled, and thus must again defer awakening its ->nocb_gp_kthread. However, ->nocb_defer_wakeup prevents CPU 0 from queueing the timer.
I managed to recollect some pieces of my brain. So keep the above but let's change the point 10:
10. CPU 0 enqueues its second callback, this time with interrupts enabled so it can wake directly ->nocb_gp_kthread. It does so with calling __wake_nocb_gp() which also cancels the pending timer that got queued in step 2. But that doesn't reset CPU 0's ->nocb_defer_wakeup which is still set to RCU_NOCB_WAKE. So CPU 0's ->nocb_defer_wakeup and CPU 0's ->nocb_timer are now desynchronized.
11. ->nocb_gp_kthread associates the callback queued in 10 with a new grace period, arrange for it to start and sleeps on it.
12. The grace period ends, ->nocb_gp_kthread awakens and wakes up CPU 0's ->nocb_cb_kthread which invokes the callback queued in 10.
13. CPU 0 enqueues its third callback, this time with interrupts disabled so it tries to queue a deferred wakeup. However ->nocb_defer_wakeup has a stalled RCU_NOCB_WAKE value which prevents the CPU 0's ->nocb_timer, that got cancelled in 10, from being armed.
14. CPU 0 has its pending callback and it may go unnoticed until some other CPU ever wakes up ->nocb_gp_kthread or CPU 0 ever calls an explicit deferred wake up caller like idle entry.
I hope I'm not missing something this time...
Thanks.
So far so good, but why isn't the timer still queued from back in step 2? What am I missing here? Either way, could you please update the commit logs to tell the full story? At some later time, you might be very happy that you did. ;-)
Thanx, Paul
- The "nocb_bypass_timer" ends up calling wake_nocb_gp() which deletes the pending "nocb_timer" (note they are not the same timers) for the given rdp without resetting the matching state stored in nocb_defer wakeup.
On both situations, a future call_rcu() on that rdp may be fooled and think the timer is armed when it's not, missing a deferred nocb_gp wakeup.
Case 1) is very unlikely due to timing constraint (the timer fires after 1 jiffy) but still possible in theory. Case 2) is more likely to happen. But in any case such scenario require the CPU to spend a long time within a kernel thread without exiting to idle or user space, which is a pretty exotic behaviour.
Fix this with resetting rdp->nocb_defer_wakeup everytime we disarm the timer.
Fixes: d1b222c6be1f (rcu/nocb: Add bypass callback queueing) Cc: Stable stable@vger.kernel.org Cc: Josh Triplett josh@joshtriplett.org Cc: Lai Jiangshan jiangshanlai@gmail.com Cc: Joel Fernandes joel@joelfernandes.org Cc: Neeraj Upadhyay neeraju@codeaurora.org Cc: Boqun Feng boqun.feng@gmail.com Signed-off-by: Frederic Weisbecker frederic@kernel.org
kernel/rcu/tree_plugin.h | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h index 2ec9d7f55f99..dd0dc66c282d 100644 --- a/kernel/rcu/tree_plugin.h +++ b/kernel/rcu/tree_plugin.h @@ -1720,7 +1720,11 @@ static bool wake_nocb_gp(struct rcu_data *rdp, bool force, rcu_nocb_unlock_irqrestore(rdp, flags); return false; }
- del_timer(&rdp->nocb_timer);
- if (READ_ONCE(rdp->nocb_defer_wakeup) > RCU_NOCB_WAKE_NOT) {
WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);del_timer(&rdp->nocb_timer);- } rcu_nocb_unlock_irqrestore(rdp, flags); raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags); if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
@@ -2349,7 +2353,6 @@ static bool do_nocb_deferred_wakeup_common(struct rcu_data *rdp) return false; } ndw = READ_ONCE(rdp->nocb_defer_wakeup);
- WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT); ret = wake_nocb_gp(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags); trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2.25.1