Re: [patch 3/3] x86: kvm guest side support for KVM_HC_RT_PRIO hypercall

From: Peter Zijlstra
Date: Thu Sep 21 2017 - 10:06:40 EST


On Thu, Sep 21, 2017 at 09:36:53AM -0400, Konrad Rzeszutek Wilk wrote:
> On Thu, Sep 21, 2017 at 08:38:38AM -0300, Marcelo Tosatti wrote:
> > Add hypercalls to spinlock/unlock to set/unset FIFO priority
> > for the vcpu, protected by a static branch to avoid performance
> > increase in the normal kernels.
> >
> > Enable option by "kvmfifohc" kernel command line parameter (disabled
> > by default).

WTF kind of fudge is this? Changelog completely fails to explain the
problem this would solve. Why are you doing insane things like this?


NAK!

> > Index: kvm.fifopriohc-submit/include/linux/spinlock_api_smp.h
> > ===================================================================
> > --- kvm.fifopriohc-submit.orig/include/linux/spinlock_api_smp.h
> > +++ kvm.fifopriohc-submit/include/linux/spinlock_api_smp.h
> > @@ -136,11 +136,28 @@ static inline void __raw_spin_lock_bh(ra
> > LOCK_CONTENDED(lock, do_raw_spin_trylock, do_raw_spin_lock);
> > }
> >
> > +#ifdef CONFIG_KVM_GUEST
> > +DECLARE_STATIC_KEY_FALSE(kvm_fifo_hc_key);
> > +#endif
> > +
> > static inline void __raw_spin_lock(raw_spinlock_t *lock)
> > {
> > preempt_disable();
> > +
> > +#if defined(CONFIG_KVM_GUEST) && defined(CONFIG_SMP)
> > + /* enable FIFO priority */
> > + if (static_branch_unlikely(&kvm_fifo_hc_key))
> > + kvm_hypercall1(KVM_HC_RT_PRIO, 0x1);
> > +#endif
> > +
> > spin_acquire(&lock->dep_map, 0, 0, _RET_IP_);
> > LOCK_CONTENDED(lock, do_raw_spin_trylock, do_raw_spin_lock);
> > +
> > +#if defined(CONFIG_KVM_GUEST) && defined(CONFIG_SMP)
> > + /* disable FIFO priority */
> > + if (static_branch_unlikely(&kvm_fifo_hc_key))
> > + kvm_hypercall1(KVM_HC_RT_PRIO, 0);
> > +#endif
> > }
> >
> > #endif /* !CONFIG_GENERIC_LOCKBREAK || CONFIG_DEBUG_LOCK_ALLOC */