Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Qiuxu Zhuo <qiuxu.zhuo@intel.com>
To:     Peter Zijlstra <peterz@infradead.org>,
        Waiman Long <longman@redhat.com>,
        Ingo Molnar <mingo@redhat.com>, Will Deacon <will@kernel.org>
Cc:     Qiuxu Zhuo <qiuxu.zhuo@intel.com>,
        Boqun Feng <boqun.feng@gmail.com>, linux-kernel@vger.kernel.org
Subject: [PATCH 1/1] locking/qspinlock: Make the 1st spinner only spin on locked_pending bits
Date:   Mon,  8 May 2023 16:15:32 +0800
Message-Id: <20230508081532.36379-1-qiuxu.zhuo@intel.com>
Precedence: bulk

The 1st spinner (header of the MCS queue) spins on the whole qspinlock
variable to check whether the lock is released. For a contended qspinlock,
this spinning is a hotspot as each CPU queued in the MCS queue performs
the spinning when it becomes the 1st spinner (header of the MCS queue).

The granularity among SMT h/w threads in the same core could be "byte"
which the Load-Store Unit (LSU) inside the core handles. Making the 1st
spinner only spin on locked_pending bits (not the whole qspinlock) can
avoid the false dependency between the tail field and the locked_pending
field. So this micro-optimization helps the h/w thread (the 1st spinner)
stay in a low power state and prevents it from being woken up by other
h/w threads in the same core when they perform xchg_tail() to update
the tail field. Please see a similar discussion in the link [1].

[1] https://lore.kernel.org/r/20230105021952.3090070-1-guoren@kernel.org

Signed-off-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
---
 kernel/locking/qspinlock.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c
index efebbf19f887..e7b990b28610 100644
--- a/kernel/locking/qspinlock.c
+++ b/kernel/locking/qspinlock.c
@@ -513,7 +513,20 @@ void __lockfunc queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
 	if ((val = pv_wait_head_or_lock(lock, node)))
 		goto locked;
 
+#if _Q_PENDING_BITS == 8
+	/*
+	 * Spinning on the 2-byte locked_pending instead of the 4-byte qspinlock
+	 * variable can avoid the false dependency between the tail field and
+	 * the locked_pending field. This helps the h/w thread (the 1st spinner)
+	 * stay in a low power state and prevents it from being woken up by other
+	 * h/w threads in the same core when they perform xchg_tail() to update
+	 * the tail field only.
+	 */
+	smp_cond_load_acquire(&lock->locked_pending, !VAL);
+	val = atomic_read_acquire(&lock->val);
+#else
 	val = atomic_cond_read_acquire(&lock->val, !(VAL & _Q_LOCKED_PENDING_MASK));
+#endif
 
 locked:
 	/*
-- 
2.17.1