On 5/16/24 06:02, Chen Yu wrote: > Performance drop is reported when running encode/decode workload and > BenchSEE cache sub-workload. > Bisect points to commit ce0a1b608bfc ("x86/paravirt: Silence unused > native_pv_lock_init() function warning"). When CONFIG_PARAVIRT_SPINLOCKS > is disabled the virt_spin_lock_key is set to true on bare-metal. > The qspinlock degenerates to test-and-set spinlock, which decrease the > performance on bare-metal. > > Fix this by disabling virt_spin_lock_key if CONFIG_PARAVIRT_SPINLOCKS > is not set, or it is on bare-metal. This is missing some background: The kernel can change spinlock behavior when running as a guest. But this guest-friendly behavior causes performance problems on bare metal. So there's a 'virt_spin_lock_key' static key to switch between the two modes. The static key is always enabled by default (run in guest mode) and should be disabled for bare metal (and in some guests that want native behavior). ... then describe the regression and the fix > diff --git a/arch/x86/kernel/paravirt.c b/arch/x86/kernel/paravirt.c > index 5358d43886ad..ee51c0949ed8 100644 > --- a/arch/x86/kernel/paravirt.c > +++ b/arch/x86/kernel/paravirt.c > @@ -55,7 +55,7 @@ DEFINE_STATIC_KEY_TRUE(virt_spin_lock_key); > > void __init native_pv_lock_init(void) > { > - if (IS_ENABLED(CONFIG_PARAVIRT_SPINLOCKS) && > + if (!IS_ENABLED(CONFIG_PARAVIRT_SPINLOCKS) || > !boot_cpu_has(X86_FEATURE_HYPERVISOR)) > static_branch_disable(&virt_spin_lock_key); > } This gets used at a single site: if (pv_enabled()) goto pv_queue; if (virt_spin_lock(lock)) return; which is logically: if (IS_ENABLED(CONFIG_PARAVIRT_SPINLOCKS)) goto ...; // don't look at virt_spin_lock_key if (virt_spin_lock_key) return; // On virt, but non-paravirt. Did Test-and-Set // spinlock. So I _think_ Arnd was trying to optimize native_pv_lock_init() away when it's going to get skipped over anyway by the 'goto'. But this took me at least 30 minutes of scratching my head and trying to untangle the whole thing. It's all far too subtle for my taste, and all of that to save a few bytes of init text in a configuration that's probably not even used very often (PARAVIRT=y, but PARAVIRT_SPINLOCKS=n). Let's just keep it simple. How about the attached patch?