Received-SPF: pass (google.com: domain of linux-kernel+bounces-22830-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
Reply-To: Sean Christopherson <seanjc@google.com>
Date: Wed, 10 Jan 2024 13:47:23 -0800
Precedence: bulk
Mime-Version: 1.0
Message-ID: <20240110214723.695930-1-seanjc@google.com>
Subject: [PATCH] sched/core: Drop spinlocks on contention iff kernel is preemptible
From: Sean Christopherson <seanjc@google.com>
To: Ingo Molnar <mingo@redhat.com>, Peter Zijlstra <peterz@infradead.org>, 
	Juri Lelli <juri.lelli@redhat.com>, Vincent Guittot <vincent.guittot@linaro.org>
Cc: linux-kernel@vger.kernel.org, 
	Valentin Schneider <valentin.schneider@arm.com>, Marco Elver <elver@google.com>, 
	Frederic Weisbecker <frederic@kernel.org>, David Matlack <dmatlack@google.com>, 
	Sean Christopherson <seanjc@google.com>
Content-Type: text/plain; charset="UTF-8"

Use preempt_model_preemptible() to detect a preemptible kernel when
deciding whether or not to reschedule in order to drop a contended
spinlock or rwlock.  Because PREEMPT_DYNAMIC selects PREEMPTION, kernels
built with PREEMPT_DYNAMIC=y will yield contended locks even if the live
preemption model is "none" or "voluntary".  In short, make kernels with
dynamically selected models behave the same as kernels with statically
selected models.

Somewhat counter-intuitively, NOT yielding a lock can provide better
latency for the relevant tasks/processes.  E.g. KVM x86's mmu_lock, a
rwlock, is often contended between an invalidation event (takes mmu_lock
for write) and a vCPU servicing a guest page fault (takes mmu_lock for
read).  For _some_ setups, letting the invalidation task complete even
if there is mmu_lock contention provides lower latency for *all* tasks,
i.e. the invalidation completes sooner *and* the vCPU services the guest
page fault sooner.

But even KVM's mmu_lock behavior isn't uniform, e.g. the "best" behavior
can vary depending on the host VMM, the guest workload, the number of
vCPUs, the number of pCPUs in the host, why there is lock contention, etc.

In other words, simply deleting the CONFIG_PREEMPTION guard (or doing the
opposite and removing contention yielding entirely) needs to come with a
big pile of data proving that changing the status quo is a net positive.

Cc: Valentin Schneider <valentin.schneider@arm.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Marco Elver <elver@google.com>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: David Matlack <dmatlack@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
---
 include/linux/sched.h | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 03bfe9ab2951..51550e64ce14 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -2229,11 +2229,10 @@ static inline bool preempt_model_preemptible(void)
  */
 static inline int spin_needbreak(spinlock_t *lock)
 {
-#ifdef CONFIG_PREEMPTION
+	if (!preempt_model_preemptible())
+		return 0;
+
 	return spin_is_contended(lock);
-#else
-	return 0;
-#endif
 }
 
 /*
@@ -2246,11 +2245,10 @@ static inline int spin_needbreak(spinlock_t *lock)
  */
 static inline int rwlock_needbreak(rwlock_t *lock)
 {
-#ifdef CONFIG_PREEMPTION
+	if (!preempt_model_preemptible())
+		return 0;
+
 	return rwlock_is_contended(lock);
-#else
-	return 0;
-#endif
 }
 
 static __always_inline bool need_resched(void)

base-commit: cdb3033e191fd03da2d7da23b9cd448dfa180a8e
-- 
2.43.0.275.g3460e3d667-goog