[PATCH] locking/paravirt: Fix performance regression on core bonded vCPU
From: johnnyaiai
Date: Thu Nov 03 2022 - 08:13:43 EST
From: johnnyaiai <johnnyaiai@xxxxxxxxxxx>
virt_spin_lock() is preferred over native qspinlock when
vCPU is preempted. But brings a lot of regression while
vCPU is not preempted. Provide a early param 'novirtlock'
to choose would be better.
will-it-scale/lock2_threads -s 10 -t 64
baseline afterpatch
559938 2166135
Signed-off-by: johnnyaiai <johnnyaiai@xxxxxxxxxxx>
---
arch/x86/kernel/paravirt.c | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/arch/x86/kernel/paravirt.c b/arch/x86/kernel/paravirt.c
index 59d3d2763..529cf23fe 100644
--- a/arch/x86/kernel/paravirt.c
+++ b/arch/x86/kernel/paravirt.c
@@ -107,6 +107,13 @@ static unsigned paravirt_patch_jmp(void *insn_buff, const void *target,
DEFINE_STATIC_KEY_TRUE(virt_spin_lock_key);
+static __init int parse_novirtspin(char *arg)
+{
+ static_branch_disable(&virt_spin_lock_key);
+ return 0;
+}
+early_param("novirtspin", parse_novirtspin);
+
void __init native_pv_lock_init(void)
{
if (!boot_cpu_has(X86_FEATURE_HYPERVISOR))
--
2.27.0