On Tue, Sep 05, 2017 at 10:14:21AM +0200, Juergen Gross wrote: > On 05/09/17 10:10, Peter Zijlstra wrote: > > On Tue, Sep 05, 2017 at 09:35:40AM +0200, Juergen Gross wrote: > >>> So the problem with qspinlock is that it will revert to a classic > >>> test-and-set spinlock if you don't do paravirt but are running a HV. > >> > >> In the Xen case we just use the bare metal settings when xen_nopvspin > >> has been specified. So paravirt, but without modifying any pv_lock_ops > >> functions. > > > > See arch/x86/include/asm/qspinlock.h:virt_spin_lock(). Unless you clear > > X86_FEATURE_HYPERVISOR you get a test-and-set spinlock. > > > > And as the comment there says, this is a fallback for !paravirt enabled > > hypervisors to avoid the worst of the lock holder preemption crud. > > > > But this very much does not deal with the 1:1 case nicely. > > > > Aah, now I've got it. > > So maybe we should add virt_spin_lock() to pv_lock_ops? This way e.g. > xen_nopvspin could tweak just the virt_spin_lock() case by letting it > return false all the time? Hmm, that might work. Could we somehow nop that call when !X86_FEATURE_HYPERVISOR?, that saves native from having to do the call and would be a win for everyone.
