rcu: Add comments to final rcu_gp_cleanup() "if" statement
[linux-2.6-microblaze.git] / kernel / rcu / tree.c
index a5ea674..2966907 100644 (file)
@@ -2098,14 +2098,29 @@ static noinline void rcu_gp_cleanup(void)
        /* Advance CBs to reduce false positives below. */
        offloaded = rcu_rdp_is_offloaded(rdp);
        if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
+
+               // We get here if a grace period was needed (“needgp”)
+               // and the above call to rcu_accelerate_cbs() did not set
+               // the RCU_GP_FLAG_INIT bit in ->gp_state (which records
+               // the need for another grace period).  The purpose
+               // of the “offloaded” check is to avoid invoking
+               // rcu_accelerate_cbs() on an offloaded CPU because we do not
+               // hold the ->nocb_lock needed to safely access an offloaded
+               // ->cblist.  We do not want to acquire that lock because
+               // it can be heavily contended during callback floods.
+
                WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
                WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
-               trace_rcu_grace_period(rcu_state.name,
-                                      rcu_state.gp_seq,
-                                      TPS("newreq"));
+               trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("newreq"));
        } else {
-               WRITE_ONCE(rcu_state.gp_flags,
-                          rcu_state.gp_flags & RCU_GP_FLAG_INIT);
+
+               // We get here either if there is no need for an
+               // additional grace period or if rcu_accelerate_cbs() has
+               // already set the RCU_GP_FLAG_INIT bit in ->gp_flags. 
+               // So all we need to do is to clear all of the other
+               // ->gp_flags bits.
+
+               WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & RCU_GP_FLAG_INIT);
        }
        raw_spin_unlock_irq_rcu_node(rnp);