On Tue, 2019-06-18 at 10:35 +0000, Roman Kagan wrote:
> On Tue, Jun 18, 2019 at 11:24:57AM +1000, Vadim Rozenfeld wrote:
> > On Mon, 2019-06-17 at 14:49 -0300, Eduardo Habkost wrote:
> > > On Mon, Jun 17, 2019 at 05:32:13PM +0000, Roman Kagan wrote:
> > > > On Mon, Jun 17, 2019 at 11:23:01AM -0300, Eduardo Habkost
> > > > wrote:
> > > > > On Mon, Jun 17, 2019 at 01:48:59PM +0000, Roman Kagan wrote:
> > > > > > On Sat, Jun 15, 2019 at 05:05:05PM -0300, Eduardo Habkost
> > > > > > wrote:
> > > > > > > The current default value for hv-spinlocks is 0xFFFFFFFF
> > > > > > > (meaning
> > > > > > > "never retry"). However, the value is stored as a signed
> > > > > > > integer, making the getter of the hv-spinlocks QOM
> > > > > > > property
> > > > > > > return -1 instead of 0xFFFFFFFF.
> > > > > > >
> > > > > > > Fix this by changing the type of
> > > > > > > X86CPU::hyperv_spinlock_attempts
> > > > > > > to uint32_t. This has no visible effect to guest
> > > > > > > operating
> > > > > > > systems, affecting just the behavior of the QOM getter.
> > > > > > >
> > > > > > > Signed-off-by: Eduardo Habkost <ehabk...@redhat.com>
> > > > > > > ---
> > > > > > > target/i386/cpu.h | 2 +-
> > > > > > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > > > > >
> > > > > > Reviewed-by: Roman Kagan <rka...@virtuozzo.com>
> > > > > >
> > > > > > That said, it's tempting to just nuke qdev_prop_spinlocks
> > > > > > and
> > > > > > make
> > > > > > hv-spinlocks a regular DEFINE_PROP_UINT32...
> > > > >
> > > > > Agreed. The only difference is that we would validate the
> > > > > property at realize time instead of object_property_set().
> > > >
> > > > Right. But currently it's validated to be no less than 0xfff
> > > > and
> > > > no
> > > > bigger than 0xffffffff. The latter check would become
> > > > unnecessary,
> > > > and
> > > > I'm unable to find any reason to do the former (neither spec
> > > > references
> > > > nor the log messages of the commits that introduced it).
> > >
> > > The 0xFFF lower limit was originally introduced by commit
> > > 28f52cc04d34 ("hyper-v: introduce Hyper-V support
> > > infrastructure").
> > >
> > > Vadim, do you know where the 0xFFF limit comes from?
> >
> > I simply took this value from Windows Server 2008 R2 that
> > I used as a reference while working on Hyper-V support for KVM.
> > I also remember some paper (probably published by AMD ???)
> > mentioned
> > that 0x2fff seemed to have the best balance for PLE logic.
>
> The question is whether the user should be disallowed to set it below
> 0xfff?
> I don't see this mandated by the spec, so I'd rather remove the lower
> limit and convert the property to a regular DEFINE_PROP_UINT32.
>
Honestly, I don't have any strong opinions on this matter. Having some
lower boundary limit seemed quite logical to me. However, if a user
wants to experiment and see how the smaller number of spinlock acquire
attempts before calling HvNotifyLongSpinWait will affect the overall
system performance, then why not?
Vadim.
> Roman.
