sched: improve sched_clock() performance
in scheduler-intense workloads native_read_tsc() overhead accounts for
20% of the system overhead:
659567 system_call 41222.9375
686796 schedule 435.7843
718382 __switch_to 665.1685
823875 switch_mm 4526.7857
1883122 native_read_tsc 55385.9412
9761990 total 2.8468
this is large part due to the rdtsc_barrier() that is done before
and after reading the TSC.
But sched_clock() is not a precise clock in the GTOD sense, using such
barriers is completely pointless. So remove the barriers and only use
them in vget_cycles().
This improves lat_ctx performance by about 5%.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
diff --git a/arch/x86/include/asm/tsc.h b/arch/x86/include/asm/tsc.h
index 38ae163..9cd83a8 100644
--- a/arch/x86/include/asm/tsc.h
+++ b/arch/x86/include/asm/tsc.h
@@ -34,6 +34,8 @@
static __always_inline cycles_t vget_cycles(void)
{
+ cycles_t cycles;
+
/*
* We only do VDSOs on TSC capable CPUs, so this shouldnt
* access boot_cpu_data (which is not VDSO-safe):
@@ -42,7 +44,11 @@
if (!cpu_has_tsc)
return 0;
#endif
- return (cycles_t)__native_read_tsc();
+ rdtsc_barrier();
+ cycles = (cycles_t)__native_read_tsc();
+ rdtsc_barrier();
+
+ return cycles;
}
extern void tsc_init(void);