| Message ID | 81da85eb-2e8e-9b76-2fb3-2beddc33e9af@suse.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | x86/time: calibration rendezvous adjustments | expand |
On Thu, Apr 01, 2021 at 11:54:27AM +0200, Jan Beulich wrote: > Since we'd like the updates to be done as synchronously as possible, > make an attempt at yielding immediately after the TSC write. > > Signed-off-by: Jan Beulich <jbeulich@suse.com> Reviewed-by: Roger Pau Monné <roger.pau@citrix.com> Did you observe any difference with the pause inserted? I wonder whether that's enough to give a chance the hyperthread to also perform the TSC write. In any case there's no harm from it certainly. Thanks, Roger.
On 20.04.2021 17:59, Roger Pau Monné wrote: > On Thu, Apr 01, 2021 at 11:54:27AM +0200, Jan Beulich wrote: >> Since we'd like the updates to be done as synchronously as possible, >> make an attempt at yielding immediately after the TSC write. >> >> Signed-off-by: Jan Beulich <jbeulich@suse.com> > > Reviewed-by: Roger Pau Monné <roger.pau@citrix.com> Thanks. > Did you observe any difference with the pause inserted? I wouldn't even know how to measure it precisely enough. So - no, I haven't. In fact ... > I wonder whether that's enough to give a chance the hyperthread to > also perform the TSC write. In any case there's no harm from it > certainly. ... I have an inquiry pending with Intel as to better (more reliable) ways to yield, for quite a bit longer than the 8259 one, yet with the same lack of any outcome so far. Until then it really is going to be no more than "make an attempt", as said in the commit message. Jan
--- a/xen/arch/x86/time.c +++ b/xen/arch/x86/time.c @@ -1708,7 +1708,14 @@ static void time_calibration_tsc_rendezv atomic_inc(&r->semaphore); if ( i == 0 ) + { write_tsc(master_tsc); + /* + * Try to give our hyperthread(s), if any, a chance to do + * the same as instantly as possible. + */ + cpu_relax(); + } while ( atomic_read(&r->semaphore) != (2*total_cpus - 1) ) cpu_relax(); @@ -1730,7 +1737,14 @@ static void time_calibration_tsc_rendezv } if ( i == 0 ) + { write_tsc(master_tsc); + /* + * Try to give our hyperthread(s), if any, a chance to do + * the same as instantly as possible. + */ + cpu_relax(); + } atomic_inc(&r->semaphore); while ( atomic_read(&r->semaphore) > total_cpus )
Since we'd like the updates to be done as synchronously as possible, make an attempt at yielding immediately after the TSC write. Signed-off-by: Jan Beulich <jbeulich@suse.com> --- v4: New.