[v5] Add HARDLOCKUP_DETECTOR_PERF support for RISC-V

[PATCH v5 0/2] Add HARDLOCKUP_DETECTOR_PERF support for RISC-V

Posted by Yunhui Cui 2 months, 4 weeks ago

After discussion [1],[2], hardlockup_perf cannot be completely replaced by
the hardlockup_buddy approach, so this patch is still being submitted.

v1->v2:
The contents of arch/arm64/watchdog_hld.c are directly consolidated into
kernel/watchdog_perf.c.


v2->v3:
Add CONFIG_WATCHDOG_PERF_ADJUST_PERIOD to enclose the period update logic,
select it by default on arm64 and riscv, without affecting other arches
like x86 and PPC.

v3->v4:
Place the line "select WATCHDOG_PERF_ADJUST_PERIOD if HARDLOCKUP_DETECTOR_PERF && CPU_FREQ"
in the Kconfig files for arm64 and riscv in one line, with no line breaks.

v4->v5:
Remove __weak from hw_nmi_get_sample_period()

Link: https://lore.kernel.org/all/CAD=FV=UEhVCD6JehQi1wor2sSmtTLDyf=37xfo-DOTK1=u1xzA@mail.gmail.com/ [1]
Link: https://lore.kernel.org/all/20250916145122.416128-1-wangjinchao600@gmail.com/ [2]

Yunhui Cui (2):
  watchdog: move arm64 watchdog_hld into common code
  riscv: add HARDLOCKUP_DETECTOR_PERF support

 arch/arm64/Kconfig               |  1 +
 arch/arm64/kernel/Makefile       |  1 -
 arch/arm64/kernel/watchdog_hld.c | 94 --------------------------------
 arch/riscv/Kconfig               |  3 +
 drivers/perf/arm_pmu.c           | 10 +++-
 drivers/perf/riscv_pmu_sbi.c     | 10 ++++
 include/linux/perf/arm_pmu.h     |  2 -
 kernel/watchdog_perf.c           | 83 ++++++++++++++++++++++++++++
 lib/Kconfig.debug                |  8 +++
 9 files changed, 114 insertions(+), 98 deletions(-)
 delete mode 100644 arch/arm64/kernel/watchdog_hld.c

-- 
2.39.5

Re: [PATCH v5 0/2] Add HARDLOCKUP_DETECTOR_PERF support for RISC-V

Posted by Doug Anderson 2 months, 3 weeks ago

Hi,

On Mon, Nov 10, 2025 at 11:35 PM Yunhui Cui <cuiyunhui@bytedance.com> wrote:
>
> After discussion [1],[2], hardlockup_perf cannot be completely replaced by
> the hardlockup_buddy approach, so this patch is still being submitted.
>
> v1->v2:
> The contents of arch/arm64/watchdog_hld.c are directly consolidated into
> kernel/watchdog_perf.c.
>
>
> v2->v3:
> Add CONFIG_WATCHDOG_PERF_ADJUST_PERIOD to enclose the period update logic,
> select it by default on arm64 and riscv, without affecting other arches
> like x86 and PPC.
>
> v3->v4:
> Place the line "select WATCHDOG_PERF_ADJUST_PERIOD if HARDLOCKUP_DETECTOR_PERF && CPU_FREQ"
> in the Kconfig files for arm64 and riscv in one line, with no line breaks.
>
> v4->v5:
> Remove __weak from hw_nmi_get_sample_period()
>
> Link: https://lore.kernel.org/all/CAD=FV=UEhVCD6JehQi1wor2sSmtTLDyf=37xfo-DOTK1=u1xzA@mail.gmail.com/ [1]
> Link: https://lore.kernel.org/all/20250916145122.416128-1-wangjinchao600@gmail.com/ [2]
>
> Yunhui Cui (2):
>   watchdog: move arm64 watchdog_hld into common code
>   riscv: add HARDLOCKUP_DETECTOR_PERF support
>
>  arch/arm64/Kconfig               |  1 +
>  arch/arm64/kernel/Makefile       |  1 -
>  arch/arm64/kernel/watchdog_hld.c | 94 --------------------------------
>  arch/riscv/Kconfig               |  3 +
>  drivers/perf/arm_pmu.c           | 10 +++-
>  drivers/perf/riscv_pmu_sbi.c     | 10 ++++
>  include/linux/perf/arm_pmu.h     |  2 -
>  kernel/watchdog_perf.c           | 83 ++++++++++++++++++++++++++++
>  lib/Kconfig.debug                |  8 +++
>  9 files changed, 114 insertions(+), 98 deletions(-)
>  delete mode 100644 arch/arm64/kernel/watchdog_hld.c

Perhaps you missed that (I think) Andrew had already dropped your v4
due to the Kernel Test Robot finding a problem. See:

https://lore.kernel.org/r/202510240701.eB6r97V3-lkp@intel.com

I don't think you've fixed these problems in v5, have you? I think you
need a v6 for it? I haven't dug all the way, but I'd suspect the
compile error is if "HARDLOCKUP_DETECTOR_PERF" is defined but not
"CPU_FREQ". Indeed, that matches the config in the report...


-Doug

Re: [External] Re: [PATCH v5 0/2] Add HARDLOCKUP_DETECTOR_PERF support for RISC-V

Posted by yunhui cui 2 months, 3 weeks ago

Hi Doug,

On Fri, Nov 14, 2025 at 6:51 AM Doug Anderson <dianders@chromium.org> wrote:
>
> Hi,
>
> On Mon, Nov 10, 2025 at 11:35 PM Yunhui Cui <cuiyunhui@bytedance.com> wrote:
> >
> > After discussion [1],[2], hardlockup_perf cannot be completely replaced by
> > the hardlockup_buddy approach, so this patch is still being submitted.
> >
> > v1->v2:
> > The contents of arch/arm64/watchdog_hld.c are directly consolidated into
> > kernel/watchdog_perf.c.
> >
> >
> > v2->v3:
> > Add CONFIG_WATCHDOG_PERF_ADJUST_PERIOD to enclose the period update logic,
> > select it by default on arm64 and riscv, without affecting other arches
> > like x86 and PPC.
> >
> > v3->v4:
> > Place the line "select WATCHDOG_PERF_ADJUST_PERIOD if HARDLOCKUP_DETECTOR_PERF && CPU_FREQ"
> > in the Kconfig files for arm64 and riscv in one line, with no line breaks.
> >
> > v4->v5:
> > Remove __weak from hw_nmi_get_sample_period()
> >
> > Link: https://lore.kernel.org/all/CAD=FV=UEhVCD6JehQi1wor2sSmtTLDyf=37xfo-DOTK1=u1xzA@mail.gmail.com/ [1]
> > Link: https://lore.kernel.org/all/20250916145122.416128-1-wangjinchao600@gmail.com/ [2]
> >
> > Yunhui Cui (2):
> >   watchdog: move arm64 watchdog_hld into common code
> >   riscv: add HARDLOCKUP_DETECTOR_PERF support
> >
> >  arch/arm64/Kconfig               |  1 +
> >  arch/arm64/kernel/Makefile       |  1 -
> >  arch/arm64/kernel/watchdog_hld.c | 94 --------------------------------
> >  arch/riscv/Kconfig               |  3 +
> >  drivers/perf/arm_pmu.c           | 10 +++-
> >  drivers/perf/riscv_pmu_sbi.c     | 10 ++++
> >  include/linux/perf/arm_pmu.h     |  2 -
> >  kernel/watchdog_perf.c           | 83 ++++++++++++++++++++++++++++
> >  lib/Kconfig.debug                |  8 +++
> >  9 files changed, 114 insertions(+), 98 deletions(-)
> >  delete mode 100644 arch/arm64/kernel/watchdog_hld.c
>
> Perhaps you missed that (I think) Andrew had already dropped your v4
> due to the Kernel Test Robot finding a problem. See:
>
> https://lore.kernel.org/r/202510240701.eB6r97V3-lkp@intel.com
>
> I don't think you've fixed these problems in v5, have you? I think you
> need a v6 for it? I haven't dug all the way, but I'd suspect the
> compile error is if "HARDLOCKUP_DETECTOR_PERF" is defined but not
> "CPU_FREQ". Indeed, that matches the config in the report...
>

Thanks for the reminder. I indeed missed replying to the email dated
October 26th regarding the error "ld.lld: error: undefined symbol:
hw_nmi_get_sample_period".
We can fix it directly as follows:
-       select WATCHDOG_PERF_ADJUST_PERIOD if HARDLOCKUP_DETECTOR_PERF
&& CPU_FREQ
+       select WATCHDOG_PERF_ADJUST_PERIOD if HARDLOCKUP_DETECTOR_PERF
Because when CPU_FREQ is not defined, there is a stub function, and
the corresponding max_cpu_freq will be set to SAFE_MAX_CPU_FREQ.

>
> -Doug

Thanks,
Yunhui