Message ID | 56d70757-a887-6824-18f4-93b1f244e44b@suse.com (mailing list archive) |
---|---|
State | Superseded |
Headers | show |
Series | x86/time: calibration rendezvous adjustments | expand |
On 09.02.2021 13:57, Jan Beulich wrote: > To reduce latency on time_calibration_tsc_rendezvous()'s last loop > iteration, separate fields written on the last iteration enough from the > crucial field read by all CPUs on the last iteration such that they end > up in distinct cache lines. Prefetch this field on an earlier iteration. I've now measured the effects of this, at least to some degree. On my single-socket 18-core Skylake system this reduces the average loop exit time on CPU0 (from the TSC write on the last iteration to until after the main loop) from around 32k cycles to around 28k (albeit the values measured on separate runs vary quite significantly). About the same effect (maybe a little less, but within error bounds) can be had without any re-arrangement of the struct's layout, by simply reading r->master_tsc_stamp into a local variable at the end of each loop iteration. I'll make v4 use this less convoluted model instead. Jan > --- a/xen/arch/x86/time.c > +++ b/xen/arch/x86/time.c > @@ -1655,10 +1655,20 @@ static void tsc_check_reliability(void) > * All CPUS snapshot their local TSC and extrapolation of system time. > */ > struct calibration_rendezvous { > - cpumask_t cpu_calibration_map; > atomic_t semaphore; > s_time_t master_stime; > - uint64_t master_tsc_stamp, max_tsc_stamp; > + cpumask_t cpu_calibration_map; > + /* > + * All CPUs want to read master_tsc_stamp on the last iteration. If > + * cpu_calibration_map isn't large enough to push the field into a cache > + * line different from the one used by semaphore (written by all CPUs on > + * every iteration) and master_stime (written by CPU0 on the last > + * iteration), align the field suitably. > + */ > + uint64_t __aligned(BITS_TO_LONGS(NR_CPUS) * sizeof(long) + > + sizeof(atomic_t) + sizeof(s_time_t) < SMP_CACHE_BYTES > + ? SMP_CACHE_BYTES : 1) master_tsc_stamp; > + uint64_t max_tsc_stamp; > }; > > static void > @@ -1709,6 +1719,8 @@ static void time_calibration_tsc_rendezv > > if ( i == 0 ) > write_tsc(r->master_tsc_stamp); > + else > + prefetch(&r->master_tsc_stamp); > > while ( atomic_read(&r->semaphore) != (2*total_cpus - 1) ) > cpu_relax(); > @@ -1731,6 +1743,8 @@ static void time_calibration_tsc_rendezv > > if ( i == 0 ) > write_tsc(r->master_tsc_stamp); > + else > + prefetch(&r->master_tsc_stamp); > > atomic_inc(&r->semaphore); > while ( atomic_read(&r->semaphore) > total_cpus ) > >
--- a/xen/arch/x86/time.c +++ b/xen/arch/x86/time.c @@ -1655,10 +1655,20 @@ static void tsc_check_reliability(void) * All CPUS snapshot their local TSC and extrapolation of system time. */ struct calibration_rendezvous { - cpumask_t cpu_calibration_map; atomic_t semaphore; s_time_t master_stime; - uint64_t master_tsc_stamp, max_tsc_stamp; + cpumask_t cpu_calibration_map; + /* + * All CPUs want to read master_tsc_stamp on the last iteration. If + * cpu_calibration_map isn't large enough to push the field into a cache + * line different from the one used by semaphore (written by all CPUs on + * every iteration) and master_stime (written by CPU0 on the last + * iteration), align the field suitably. + */ + uint64_t __aligned(BITS_TO_LONGS(NR_CPUS) * sizeof(long) + + sizeof(atomic_t) + sizeof(s_time_t) < SMP_CACHE_BYTES + ? SMP_CACHE_BYTES : 1) master_tsc_stamp; + uint64_t max_tsc_stamp; }; static void @@ -1709,6 +1719,8 @@ static void time_calibration_tsc_rendezv if ( i == 0 ) write_tsc(r->master_tsc_stamp); + else + prefetch(&r->master_tsc_stamp); while ( atomic_read(&r->semaphore) != (2*total_cpus - 1) ) cpu_relax(); @@ -1731,6 +1743,8 @@ static void time_calibration_tsc_rendezv if ( i == 0 ) write_tsc(r->master_tsc_stamp); + else + prefetch(&r->master_tsc_stamp); atomic_inc(&r->semaphore); while ( atomic_read(&r->semaphore) > total_cpus )
To reduce latency on time_calibration_tsc_rendezvous()'s last loop iteration, separate fields written on the last iteration enough from the crucial field read by all CPUs on the last iteration such that they end up in distinct cache lines. Prefetch this field on an earlier iteration. Signed-off-by: Jan Beulich <jbeulich@suse.com> --- v3: New. --- While readability would likely suffer, we may want to consider avoiding the prefetch on at least the first two iterations (where the field still gets / may get written to by CPU0). Could e.g. be switch ( i ) { case 0: write_tsc(r->master_tsc_stamp); break; case 1: prefetch(&r->master_tsc_stamp); break; } Of course it would also be nice to avoid the pretty likely branch misprediction on the last iteration. But with the static prediction hints having been rather short-lived in the architecture, I don't see any good means to do so.