vgettimeofday disabled?

vgettimeofday disabled?

[Chris Lalancette]

Hello,
     Based on a comment from a customer, I looked briefly at the vgettimeofday
implementation in Xen.  I noticed that vgettimeofday is disabled in the 64-bit
linux-2.6.18-xen.hg tree, basically by setting sysctl_vsyscall = 0 in
vsyscall_init().  Is there some known problem with enabling this?  In a quick
test, I turned it on, and went from 2m14s in my micro benchmark to 30s for the
benchmark, but I'm not sure if there will be any ill effects from the same.

Thanks,
Chris Lalancette

Re: vgettimeofday disabled?

[Keir Fraser]

On 3/9/08 16:48, "Chris Lalancette" <clalance@redhat.com> wrote:

> Hello,
>      Based on a comment from a customer, I looked briefly at the vgettimeofday
> implementation in Xen.  I noticed that vgettimeofday is disabled in the 64-bit
> linux-2.6.18-xen.hg tree, basically by setting sysctl_vsyscall = 0 in
> vsyscall_init().  Is there some known problem with enabling this?  In a quick
> test, I turned it on, and went from 2m14s in my micro benchmark to 30s for the
> benchmark, but I'm not sure if there will be any ill effects from the same.

We probably don't update vxtime structure, so you likely get dodgy results
from do_vgettimedofday(). However, forcing vxtime to track vcpu0's time info
obtained from Xen would probably be quite easy to do and may be accurate
enough.

But, without that extra engineering, I recommend you check what answers
you're getting from gettimeofday(). Could be quite a significant
performance/correctness tradeoff ;-)

 -- Keir

Re: vgettimeofday disabled?

[Chris Lalancette]

Keir Fraser wrote:
> We probably don't update vxtime structure, so you likely get dodgy results
> from do_vgettimedofday(). However, forcing vxtime to track vcpu0's time info
> obtained from Xen would probably be quite easy to do and may be accurate
> enough.
> 
> But, without that extra engineering, I recommend you check what answers
> you're getting from gettimeofday(). Could be quite a significant
> performance/correctness tradeoff ;-)

Ah, OK.  So there's not really a fundamental issue (which I was worried about),
more just it hasn't been implemented all the way.  Interestingly, it doesn't
seem to be explicitly disabled in the LKML pv_ops implementation, so either it's
being updated properly there, or time is *way* off.

Thanks for the info.

Chris Lalancette