Maxim Levitsky <mlevitsk@xxxxxxxxxx> writes: > Just something I noticed today. Happens on both AMD and Intel, kvm/queue. > > Likely the test needs lower tolerancies. > > I'll investigate this later > > This is on my AMD machine (3970X): > > [mlevitsk@starship ~/Kernel/master/src/tools/testing/selftests/kvm]$while true ; do ./x86_64/hyperv_clock ; done > ==== Test Assertion Failure ==== > x86_64/hyperv_clock.c:199: delta_ns * 100 < (t2 - t1) * 100 > pid=66218 tid=66218 errno=0 - Success > 1 0x000000000040255d: host_check_tsc_msr_rdtsc at hyperv_clock.c:199 > 2 (inlined by) main at hyperv_clock.c:223 > 3 0x00007f0f2822d55f: ?? ??:0 > 4 0x00007f0f2822d60b: ?? ??:0 > 5 0x0000000000402744: _start at ??:? > Elapsed time does not match (MSR=471600, TSC=461024) ... Here the test is: r1 = rdtsc() m1 = KVM_GET_MSRS (HV_X64_MSR_TIME_REF_COUNT) nop_loop() r2 = rdtsc() m2 = KVM_GET_MSRS (HV_X64_MSR_TIME_REF_COUNT) and then we compare the difference between rdtsc()-s and HV_X64_MSR_TIME_REF_COUNT changes with 1% tolerance (r2-r1 vs m2-m1). It would probably increase accuracy if we do r1_1 = rdtsc() KVM_GET_MSRS (HV_X64_MSR_TIME_REF_COUNT) r1_2 = rdtsc() nop_loop() r2_1 = rdtsc() KVM_GET_MSRS (HV_X64_MSR_TIME_REF_COUNT) r2_2 = rdtsc() and compare (r2_2 + r2_1)/2 - (r2_1 + r2_2)/2 vs m2-m1. and also increase tolerance to say 5%. > ==== Test Assertion Failure ==== > x86_64/hyperv_clock.c:234: false > pid=66652 tid=66652 errno=4 - Interrupted system call > 1 0x00000000004026e7: main at hyperv_clock.c:234 > 2 0x00007fdab782d55f: ?? ??:0 > 3 0x00007fdab782d60b: ?? ??:0 > 4 0x0000000000402744: _start at ??:? > Failed guest assert: delta_ns * 100 < (t2 - t1) * 100 at x86_64/hyperv_clock.c:74 Same story as above but from within the guest (rdmsr() istead of KVM_GET_MSRS). We can probably employ the same idea to increate the accuracy. -- Vitaly
