| Message ID | 20210623080157.26424-1-krzysztof.kozlowski@canonical.com |
|---|---|
| State | Accepted |
| Headers | show |
| Series | device-drivers/cpufreq_boost: skip test on virtual machines | expand |
On 23/06/2021 10:01, Krzysztof Kozlowski wrote: > Testing CPU frequency boost on a virtual machines is not reliable. Even > if cpufreq driver reports boost frequencies, the underlying hypervisor > might decide otherwise. > > For example on AWS x1e.xlarge instance (claimed to be running on Intel > Xeon E7 8880) under Xen hypervisor the test always fails: > > cpufreq_boost 0 TINFO : found 'intel_pstate' driver, sysfs knob '/sys/devices/system/cpu/intel_pstate/no_turbo' > cpufreq_boost 0 TINFO : maximum speed is 3100000 KHz > cpufreq_boost 0 TINFO : load CPU0 with boost enabled > cpufreq_boost 0 TINFO : elapsed time is 1155 ms > cpufreq_boost 0 TINFO : load CPU0 with boost disabled > cpufreq_boost 0 TINFO : elapsed time is 1155 ms > cpufreq_boost 1 TFAIL : cpufreq_boost.c:186: compare time spent with and without boost (-2%) > > Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@canonical.com> > --- > testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c b/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c > index b9739db37cb7..67917b3fea25 100644 > --- a/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c > +++ b/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c > @@ -90,6 +90,9 @@ static void setup(void) > unsigned int i; > tst_require_root(); > > + if (tst_is_virt(VIRT_ANY)) > + tst_brkm(TCONF, NULL, "running in a virtual machine, overclock not reliably measureable"); > + > for (i = 0; i < ARRAY_SIZE(cdrv); ++i) { > fd = open(cdrv[i].file, O_RDWR); > if (fd == -1) > Optionally, under virtual machine the test failure could be converted to accepted pass. This would still allow to test CPUfreq boosting interface. Any preferences? Best regards, Krzysztof
Hi! > > diff --git a/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c b/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c > > index b9739db37cb7..67917b3fea25 100644 > > --- a/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c > > +++ b/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c > > @@ -90,6 +90,9 @@ static void setup(void) > > unsigned int i; > > tst_require_root(); > > > > + if (tst_is_virt(VIRT_ANY)) > > + tst_brkm(TCONF, NULL, "running in a virtual machine, overclock not reliably measureable"); > > + > > for (i = 0; i < ARRAY_SIZE(cdrv); ++i) { > > fd = open(cdrv[i].file, O_RDWR); > > if (fd == -1) > > > > Optionally, under virtual machine the test failure could be converted to > accepted pass. This would still allow to test CPUfreq boosting > interface. Any preferences? I wonder what is the likehood of actually dicovering a bug by writing to the cpufreq boost file from within a VM, I guess that it's non-zero at least.
On 23/06/2021 13:38, Cyril Hrubis wrote: > Hi! >>> diff --git a/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c b/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c >>> index b9739db37cb7..67917b3fea25 100644 >>> --- a/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c >>> +++ b/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c >>> @@ -90,6 +90,9 @@ static void setup(void) >>> unsigned int i; >>> tst_require_root(); >>> >>> + if (tst_is_virt(VIRT_ANY)) >>> + tst_brkm(TCONF, NULL, "running in a virtual machine, overclock not reliably measureable"); >>> + >>> for (i = 0; i < ARRAY_SIZE(cdrv); ++i) { >>> fd = open(cdrv[i].file, O_RDWR); >>> if (fd == -1) >>> >> >> Optionally, under virtual machine the test failure could be converted to >> accepted pass. This would still allow to test CPUfreq boosting >> interface. Any preferences? > > I wonder what is the likehood of actually dicovering a bug by writing to > the cpufreq boost file from within a VM, I guess that it's non-zero at > least. It's a test of kernel interface and whether the driver accepts input via sysfs... so not much of a real cpufreq test. Indeed maybe better to skip the test explicitly instead of giving some false hopes that cpufreq is being tested. Best regards, Krzysztof
On 23/06/2021 17:33, Krzysztof Kozlowski wrote: > On 23/06/2021 13:38, Cyril Hrubis wrote: >> Hi! >>>> diff --git a/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c b/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c >>>> index b9739db37cb7..67917b3fea25 100644 >>>> --- a/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c >>>> +++ b/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c >>>> @@ -90,6 +90,9 @@ static void setup(void) >>>> unsigned int i; >>>> tst_require_root(); >>>> >>>> + if (tst_is_virt(VIRT_ANY)) >>>> + tst_brkm(TCONF, NULL, "running in a virtual machine, overclock not reliably measureable"); >>>> + >>>> for (i = 0; i < ARRAY_SIZE(cdrv); ++i) { >>>> fd = open(cdrv[i].file, O_RDWR); >>>> if (fd == -1) >>>> >>> >>> Optionally, under virtual machine the test failure could be converted to >>> accepted pass. This would still allow to test CPUfreq boosting >>> interface. Any preferences? >> >> I wonder what is the likehood of actually dicovering a bug by writing to >> the cpufreq boost file from within a VM, I guess that it's non-zero at >> least. > > It's a test of kernel interface and whether the driver accepts input via > sysfs... so not much of a real cpufreq test. Indeed maybe better to skip > the test explicitly instead of giving some false hopes that cpufreq is > being tested. Hi Cyril and all, Anything stops this patch from being applied? Best regards, Krzysztof
Hi! > Hi Cyril and all, > > Anything stops this patch from being applied? Thanks for the reminded, pushed.
diff --git a/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c b/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c index b9739db37cb7..67917b3fea25 100644 --- a/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c +++ b/testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c @@ -90,6 +90,9 @@ static void setup(void) unsigned int i; tst_require_root(); + if (tst_is_virt(VIRT_ANY)) + tst_brkm(TCONF, NULL, "running in a virtual machine, overclock not reliably measureable"); + for (i = 0; i < ARRAY_SIZE(cdrv); ++i) { fd = open(cdrv[i].file, O_RDWR); if (fd == -1)
Testing CPU frequency boost on a virtual machines is not reliable. Even if cpufreq driver reports boost frequencies, the underlying hypervisor might decide otherwise. For example on AWS x1e.xlarge instance (claimed to be running on Intel Xeon E7 8880) under Xen hypervisor the test always fails: cpufreq_boost 0 TINFO : found 'intel_pstate' driver, sysfs knob '/sys/devices/system/cpu/intel_pstate/no_turbo' cpufreq_boost 0 TINFO : maximum speed is 3100000 KHz cpufreq_boost 0 TINFO : load CPU0 with boost enabled cpufreq_boost 0 TINFO : elapsed time is 1155 ms cpufreq_boost 0 TINFO : load CPU0 with boost disabled cpufreq_boost 0 TINFO : elapsed time is 1155 ms cpufreq_boost 1 TFAIL : cpufreq_boost.c:186: compare time spent with and without boost (-2%) Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@canonical.com> --- testcases/kernel/device-drivers/cpufreq/cpufreq_boost.c | 3 +++ 1 file changed, 3 insertions(+)