refperf: More closely synchronize reader start times

Currently, readers are awakened individually.  On most systems, this
results in significant wakeup delay from one reader to the next, which
can result in the first and last reader having sole access to the
synchronization primitive in question.  If that synchronization primitive
involves shared memory, those readers will rack up a huge number of
operations in a very short time, causing large perturbations in the
results.

This commit therefore has the readers busy-wait after being awakened,
and uses a new n_started variable to synchronize their start times.

Cc: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>

1 files changed, 5 insertions, 0 deletions

diff --git a/kernel/rcu/refperf.c b/kernel/rcu/refperf.c
index 2fd3ed1a0d0d..234bb0e84a8b 100644
--- a/kernel/rcu/refperf.c
+++ b/kernel/rcu/refperf.c
@@ -99,6 +99,7 @@ static atomic_t nreaders_exp;
 
 // Use to wait for all threads to start.
 static atomic_t n_init;
+static atomic_t n_started;
 
 // Track which experiment is currently running.
 static int exp_idx;
@@ -253,6 +254,9 @@ repeat:
 	WARN_ON_ONCE(smp_processor_id() != me);
 
 	WRITE_ONCE(rt->start_reader, 0);
+	if (!atomic_dec_return(&n_started))
+		while (atomic_read_acquire(&n_started))
+			cpu_relax();
 
 	VERBOSE_PERFOUT("ref_perf_reader %ld: experiment %d started", me, exp_idx);
 
@@ -367,6 +371,7 @@ static int main_func(void *arg)
 
 		reset_readers();
 		atomic_set(&nreaders_exp, nreaders);
+		atomic_set(&n_started, nreaders);
 
 		exp_idx = exp;