rcu: Fix delayed execution of hurry callbacks

We observed a regression in our customer’s environment after enabling
CONFIG_LAZY_RCU. In the Android Update Engine scenario, where ioctl() is
used heavily, we found that callbacks queued via call_rcu_hurry (such as
percpu_ref_switch_to_atomic_rcu) can sometimes be delayed by up to 5
seconds before execution. This occurs because the new grace period does
not start immediately after the previous one completes.

The root cause is that the wake_nocb_gp_defer() function now checks
"rdp->nocb_defer_wakeup" instead of "rdp_gp->nocb_defer_wakeup". On CPUs
that are not rcuog, "rdp->nocb_defer_wakeup" may always be
RCU_NOCB_WAKE_NOT. This can cause "rdp_gp->nocb_defer_wakeup" to be
downgraded and the "rdp_gp->nocb_timer" to be postponed by up to 10
seconds, delaying the execution of hurry RCU callbacks.

The trace log of one scenario we encountered is as follow:
  // previous GP ends at this point
  rcu_preempt   [000] d..1.   137.240210: rcu_grace_period: rcu_preempt 8369 end
  rcu_preempt   [000] .....   137.240212: rcu_grace_period: rcu_preempt 8372 reqwait
  // call_rcu_hurry enqueues "percpu_ref_switch_to_atomic_rcu", the callback waited on by UpdateEngine
  update_engine [002] d..1.   137.301593: __call_rcu_common: wyy: unlikely p_ref = 00000000********. lazy = 0
  // FirstQ on cpu 2 rdp_gp->nocb_timer is set to fire after 1 jiffy (4ms)
  // and the rdp_gp->nocb_defer_wakeup is set to RCU_NOCB_WAKE
  update_engine [002] d..2.   137.301595: rcu_nocb_wake: rcu_preempt 2 FirstQ on cpu2 with rdp_gp (cpu0).
  // FirstBQ event on cpu2 during the 1 jiffy, make the timer postpond 10 seconds later.
  // also, the rdp_gp->nocb_defer_wakeup is overwrite to RCU_NOCB_WAKE_LAZY
  update_engine [002] d..1.   137.301601: rcu_nocb_wake: rcu_preempt 2 WakeEmptyIsDeferred
  ...
  ...
  ...
  // before the 10 seconds timeout, cpu0 received another call_rcu_hurry
  // reset the timer to jiffies+1 and set the waketype = RCU_NOCB_WAKE.
  kworker/u32:0 [000] d..2.   142.557564: rcu_nocb_wake: rcu_preempt 0 FirstQ
  kworker/u32:0 [000] d..1.   142.557576: rcu_nocb_wake: rcu_preempt 0 WakeEmptyIsDeferred
  kworker/u32:0 [000] d..1.   142.558296: rcu_nocb_wake: rcu_preempt 0 WakeNot
  kworker/u32:0 [000] d..1.   142.558562: rcu_nocb_wake: rcu_preempt 0 WakeNot
  // idle(do_nocb_deferred_wakeup) wake rcuog due to waketype == RCU_NOCB_WAKE
  <idle>        [000] d..1.   142.558786: rcu_nocb_wake: rcu_preempt 0 DoWake
  <idle>        [000] dN.1.   142.558839: rcu_nocb_wake: rcu_preempt 0 DeferredWake
  rcuog/0       [000] .....   142.558871: rcu_nocb_wake: rcu_preempt 0 EndSleep
  rcuog/0       [000] .....   142.558877: rcu_nocb_wake: rcu_preempt 0 Check
  // finally rcuog request a new GP at this point (5 seconds after the FirstQ event)
  rcuog/0       [000] d..2.   142.558886: rcu_grace_period: rcu_preempt 8372 newreq
  rcu_preempt   [001] d..1.   142.559458: rcu_grace_period: rcu_preempt 8373 start
  ...
  rcu_preempt   [000] d..1.   142.564258: rcu_grace_period: rcu_preempt 8373 end
  rcuop/2       [000] D..1.   142.566337: rcu_batch_start: rcu_preempt CBs=219 bl=10
  // the hurry CB is invoked at this point
  rcuop/2       [000] b....   142.566352: blk_queue_usage_counter_release: wyy: wakeup. p_ref = 00000000********.

This patch changes the condition to check "rdp_gp->nocb_defer_wakeup" in
the lazy path. This prevents an already scheduled "rdp_gp->nocb_timer"
from being postponed and avoids overwriting "rdp_gp->nocb_defer_wakeup"
when it is not RCU_NOCB_WAKE_NOT.

Fixes: 3cb278e ("rcu: Make call_rcu() lazy to save power")
Co-developed-by: Cheng-jui Wang <[email protected]>
Signed-off-by: Cheng-jui Wang <[email protected]>
Co-developed-by: [email protected]
Signed-off-by: [email protected]
Tested-by: [email protected]
Signed-off-by: [email protected]
Signed-off-by: Tze-nan Wu <[email protected]>
Reviewed-by: Frederic Weisbecker <[email protected]>
Signed-off-by: Neeraj Upadhyay (AMD) <[email protected]>

Author: Tze-nan Wu

kernel/rcu/tree_nocb.h

@@ -276,7 +276,7 @@ static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
 	 * callback storms, no need to wake up too early.
 	 */
 	if (waketype == RCU_NOCB_WAKE_LAZY &&
-	    rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT) {
+	    rdp_gp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT) {
 		mod_timer(&rdp_gp->nocb_timer, jiffies + rcu_get_jiffies_lazy_flush());
 		WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
 	} else if (waketype == RCU_NOCB_WAKE_BYPASS) {