| Message ID | 20240805091722.6871-1-felix.moessbauer@siemens.com (mailing list archive) |
|---|---|
| State | Accepted |
| Headers | show |
| Series | [isar-cip-core,1/1] fix(PREEMPT_RT): disable CONFIG_RT_GROUP_SCHED | expand |
On 05.08.24 11:17, Felix Moessbauer wrote: > This feature is simply broken on RT kernels that use cgroup2, as the > cpu controller features rt_period_us and rt_runtime_us are not > exposed, which are needed to configure things in a way that allows to > run tasks under RT/DL scheduling policies. It further seems to be > incompatible with the nohz_full option. > > This patch just disables the feature, similar to upstream Debian. > Makes sense. Didn't trigger yet with the standard RT test images built from isar-cip-core directly, but out config should think beyond. > Signed-off-by: Felix Moessbauer <felix.moessbauer@siemens.com> > --- > I'm wondering a bit how that was not discovered earlier in our RT tests. > Maybe Quirin can have a look if the cyclictest also shows errors that > the scheduling policy cannot be set. But maybe it simply did not show up > there, as the setup was to simplistic and did not enable the cgroup2 cpu > controller. The latter is what I assume as well. I tested it with a qemu-x86 bookworm 6.1-rt setup, and cyclictest happily worked. > > Note: This was discovered while tracing down kernel bugs around > timer slack on RT/DL tasks. > > Best regards, > Felix Moessbauer > Siemens AG > > recipes-kernel/linux/files/preempt-rt.cfg | 1 + > 1 file changed, 1 insertion(+) > > diff --git a/recipes-kernel/linux/files/preempt-rt.cfg b/recipes-kernel/linux/files/preempt-rt.cfg > index 4afe1bf..ea37556 100644 > --- a/recipes-kernel/linux/files/preempt-rt.cfg > +++ b/recipes-kernel/linux/files/preempt-rt.cfg > @@ -1,6 +1,7 @@ > # >= 5.10 > CONFIG_EXPERT=y > CONFIG_PREEMPT_RT=y > +CONFIG_RT_GROUP_SCHED=n > # <= 4.19 > CONFIG_PREEMPT_RT_FULL=y > Thanks, applied. Jan
On Mon, 2024-08-05 at 16:24 +0200, Jan Kiszka wrote: > On 05.08.24 11:17, Felix Moessbauer wrote: > > This feature is simply broken on RT kernels that use cgroup2, as > > the > > cpu controller features rt_period_us and rt_runtime_us are not > > exposed, which are needed to configure things in a way that allows > > to > > run tasks under RT/DL scheduling policies. It further seems to be > > incompatible with the nohz_full option. > > > > This patch just disables the feature, similar to upstream Debian. > > > > Makes sense. Didn't trigger yet with the standard RT test images > built > from isar-cip-core directly, but out config should think beyond. > > > Signed-off-by: Felix Moessbauer <felix.moessbauer@siemens.com> > > --- > > I'm wondering a bit how that was not discovered earlier in our RT > > tests. > > Maybe Quirin can have a look if the cyclictest also shows errors > > that > > the scheduling policy cannot be set. But maybe it simply did not > > show up > > there, as the setup was to simplistic and did not enable the > > cgroup2 cpu > > controller. > > The latter is what I assume as well. I tested it with a qemu-x86 > bookworm 6.1-rt setup, and cyclictest happily worked. I just noticed, that the CONFIG_CPU_ISOLATION is unset in the CIP RT kernel config. While this makes sense in pure cgroup2 deployments, I'm pretty sure many deployments still rely on the isolcpus= kernel parameter (which only works with CONFIG_CPU_ISOLATION=y). This makes me wonder what exactly we are testing there. Felix > > > > > Note: This was discovered while tracing down kernel bugs around > > timer slack on RT/DL tasks. > > > > Best regards, > > Felix Moessbauer > > Siemens AG > > > > recipes-kernel/linux/files/preempt-rt.cfg | 1 + > > 1 file changed, 1 insertion(+) > > > > diff --git a/recipes-kernel/linux/files/preempt-rt.cfg b/recipes- > > kernel/linux/files/preempt-rt.cfg > > index 4afe1bf..ea37556 100644 > > --- a/recipes-kernel/linux/files/preempt-rt.cfg > > +++ b/recipes-kernel/linux/files/preempt-rt.cfg > > @@ -1,6 +1,7 @@ > > # >= 5.10 > > CONFIG_EXPERT=y > > CONFIG_PREEMPT_RT=y > > +CONFIG_RT_GROUP_SCHED=n > > # <= 4.19 > > CONFIG_PREEMPT_RT_FULL=y > > > > Thanks, applied. > > Jan >
On 07.08.24 15:57, Moessbauer, Felix (T CED OES-DE) wrote: > On Mon, 2024-08-05 at 16:24 +0200, Jan Kiszka wrote: >> On 05.08.24 11:17, Felix Moessbauer wrote: >>> This feature is simply broken on RT kernels that use cgroup2, as >>> the >>> cpu controller features rt_period_us and rt_runtime_us are not >>> exposed, which are needed to configure things in a way that allows >>> to >>> run tasks under RT/DL scheduling policies. It further seems to be >>> incompatible with the nohz_full option. >>> >>> This patch just disables the feature, similar to upstream Debian. >>> >> >> Makes sense. Didn't trigger yet with the standard RT test images >> built >> from isar-cip-core directly, but out config should think beyond. >> >>> Signed-off-by: Felix Moessbauer <felix.moessbauer@siemens.com> >>> --- >>> I'm wondering a bit how that was not discovered earlier in our RT >>> tests. >>> Maybe Quirin can have a look if the cyclictest also shows errors >>> that >>> the scheduling policy cannot be set. But maybe it simply did not >>> show up >>> there, as the setup was to simplistic and did not enable the >>> cgroup2 cpu >>> controller. >> >> The latter is what I assume as well. I tested it with a qemu-x86 >> bookworm 6.1-rt setup, and cyclictest happily worked. > > I just noticed, that the CONFIG_CPU_ISOLATION is unset in the CIP RT > kernel config. While this makes sense in pure cgroup2 deployments, I'm > pretty sure many deployments still rely on the isolcpus= kernel > parameter (which only works with CONFIG_CPU_ISOLATION=y). > > This makes me wonder what exactly we are testing there. > I suppose you are referring to cip-kernel-config, and there x86/cip_merged_defconfig: We need input in form of project/device configs in order to cover what people use, to maintain and test features properly. Jan
diff --git a/recipes-kernel/linux/files/preempt-rt.cfg b/recipes-kernel/linux/files/preempt-rt.cfg index 4afe1bf..ea37556 100644 --- a/recipes-kernel/linux/files/preempt-rt.cfg +++ b/recipes-kernel/linux/files/preempt-rt.cfg @@ -1,6 +1,7 @@ # >= 5.10 CONFIG_EXPERT=y CONFIG_PREEMPT_RT=y +CONFIG_RT_GROUP_SCHED=n # <= 4.19 CONFIG_PREEMPT_RT_FULL=y
This feature is simply broken on RT kernels that use cgroup2, as the cpu controller features rt_period_us and rt_runtime_us are not exposed, which are needed to configure things in a way that allows to run tasks under RT/DL scheduling policies. It further seems to be incompatible with the nohz_full option. This patch just disables the feature, similar to upstream Debian. Signed-off-by: Felix Moessbauer <felix.moessbauer@siemens.com> --- I'm wondering a bit how that was not discovered earlier in our RT tests. Maybe Quirin can have a look if the cyclictest also shows errors that the scheduling policy cannot be set. But maybe it simply did not show up there, as the setup was to simplistic and did not enable the cgroup2 cpu controller. Note: This was discovered while tracing down kernel bugs around timer slack on RT/DL tasks. Best regards, Felix Moessbauer Siemens AG recipes-kernel/linux/files/preempt-rt.cfg | 1 + 1 file changed, 1 insertion(+)