2024-05-14 18:01:07

by Namhyung Kim

[permalink] [raw]
Subject: [PATCH] perf/arm-dmc620: Fix lockdep assert in ->event_init()

for_each_sibling_event() checks leader's ctx but it doesn't have the ctx
yet if it's the leader. Like in perf_event_validate_size(), we should
skip checking siblings in that case.

Fixes: f3c0eba287049 ("perf: Add a few assertions")
Reported-by: Greg Thelen <[email protected]>
Cc: Robin Murphy <[email protected]>
Cc: Tuan Phan <[email protected]>
Signed-off-by: Namhyung Kim <[email protected]>
---
drivers/perf/arm_dmc620_pmu.c | 9 ++++++---
1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/drivers/perf/arm_dmc620_pmu.c b/drivers/perf/arm_dmc620_pmu.c
index 8a81be2dd5ec..88c17c1d6d49 100644
--- a/drivers/perf/arm_dmc620_pmu.c
+++ b/drivers/perf/arm_dmc620_pmu.c
@@ -542,12 +542,16 @@ static int dmc620_pmu_event_init(struct perf_event *event)
if (event->cpu < 0)
return -EINVAL;

+ hwc->idx = -1;
+
+ if (event->group_leader == event)
+ return 0;
+
/*
* We can't atomically disable all HW counters so only one event allowed,
* although software events are acceptable.
*/
- if (event->group_leader != event &&
- !is_software_event(event->group_leader))
+ if (!is_software_event(event->group_leader))
return -EINVAL;

for_each_sibling_event(sibling, event->group_leader) {
@@ -556,7 +560,6 @@ static int dmc620_pmu_event_init(struct perf_event *event)
return -EINVAL;
}

- hwc->idx = -1;
return 0;
}

--
2.45.0.rc1.225.g2a3ae87e7f-goog



2024-05-15 12:12:14

by Robin Murphy

[permalink] [raw]
Subject: Re: [PATCH] perf/arm-dmc620: Fix lockdep assert in ->event_init()

On 2024-05-14 7:00 pm, Namhyung Kim wrote:
> for_each_sibling_event() checks leader's ctx but it doesn't have the ctx
> yet if it's the leader. Like in perf_event_validate_size(), we should
> skip checking siblings in that case.

Ugh, looking around for_each_sibling_event() sites, it looks like there
are a fair few other drivers using this pattern as well :(

I'd love for groups to be less horribly complicated, but I think I can
follow the underlying reasoning here. I suppose one could argue that the
assertion could take into account that there's nothing to protect in the
case where event->ctx is still NULL, since nobody else should be able to
touch the event's own empty sibling list at this point before
perf_event_open() has even returned. However by the same token there's
also no real reason for drivers *not* to return early when they equally
can tell that the sibling list must be empty, and indeed that seems to
be a fairly common pattern too, so I see no issue with fixing up all the
offending drivers for consistency either.

Reviewed-by: Robin Murphy <[email protected]>

> Fixes: f3c0eba287049 ("perf: Add a few assertions")
> Reported-by: Greg Thelen <[email protected]>
> Cc: Robin Murphy <[email protected]>
> Cc: Tuan Phan <[email protected]>
> Signed-off-by: Namhyung Kim <[email protected]>
> ---
> drivers/perf/arm_dmc620_pmu.c | 9 ++++++---
> 1 file changed, 6 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/perf/arm_dmc620_pmu.c b/drivers/perf/arm_dmc620_pmu.c
> index 8a81be2dd5ec..88c17c1d6d49 100644
> --- a/drivers/perf/arm_dmc620_pmu.c
> +++ b/drivers/perf/arm_dmc620_pmu.c
> @@ -542,12 +542,16 @@ static int dmc620_pmu_event_init(struct perf_event *event)
> if (event->cpu < 0)
> return -EINVAL;
>
> + hwc->idx = -1;
> +
> + if (event->group_leader == event)
> + return 0;
> +
> /*
> * We can't atomically disable all HW counters so only one event allowed,
> * although software events are acceptable.
> */
> - if (event->group_leader != event &&
> - !is_software_event(event->group_leader))
> + if (!is_software_event(event->group_leader))
> return -EINVAL;
>
> for_each_sibling_event(sibling, event->group_leader) {
> @@ -556,7 +560,6 @@ static int dmc620_pmu_event_init(struct perf_event *event)
> return -EINVAL;
> }
>
> - hwc->idx = -1;
> return 0;
> }
>

2024-05-17 12:05:18

by Will Deacon

[permalink] [raw]
Subject: Re: [PATCH] perf/arm-dmc620: Fix lockdep assert in ->event_init()

On Tue, May 14, 2024 at 11:00:50AM -0700, Namhyung Kim wrote:
> for_each_sibling_event() checks leader's ctx but it doesn't have the ctx
> yet if it's the leader. Like in perf_event_validate_size(), we should
> skip checking siblings in that case.
>
> Fixes: f3c0eba287049 ("perf: Add a few assertions")
> Reported-by: Greg Thelen <[email protected]>
> Cc: Robin Murphy <[email protected]>
> Cc: Tuan Phan <[email protected]>
> Signed-off-by: Namhyung Kim <[email protected]>
> ---
> drivers/perf/arm_dmc620_pmu.c | 9 ++++++---
> 1 file changed, 6 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/perf/arm_dmc620_pmu.c b/drivers/perf/arm_dmc620_pmu.c
> index 8a81be2dd5ec..88c17c1d6d49 100644
> --- a/drivers/perf/arm_dmc620_pmu.c
> +++ b/drivers/perf/arm_dmc620_pmu.c
> @@ -542,12 +542,16 @@ static int dmc620_pmu_event_init(struct perf_event *event)
> if (event->cpu < 0)
> return -EINVAL;
>
> + hwc->idx = -1;
> +
> + if (event->group_leader == event)
> + return 0;
> +
> /*
> * We can't atomically disable all HW counters so only one event allowed,
> * although software events are acceptable.
> */
> - if (event->group_leader != event &&
> - !is_software_event(event->group_leader))
> + if (!is_software_event(event->group_leader))
> return -EINVAL;
>
> for_each_sibling_event(sibling, event->group_leader) {
> @@ -556,7 +560,6 @@ static int dmc620_pmu_event_init(struct perf_event *event)
> return -EINVAL;
> }
>
> - hwc->idx = -1;
> return 0;
> }

Thanks, I'll pick this up, although Mark reckoned he'd found some other
issues over at:

https://lore.kernel.org/r/Zg0l642PgQ7T3a8Z@FVFF77S0Q05N

but didn't elaborate on what exactly he'd found :/

Will

2024-05-17 12:17:26

by Robin Murphy

[permalink] [raw]
Subject: Re: [PATCH] perf/arm-dmc620: Fix lockdep assert in ->event_init()

On 17/05/2024 1:02 pm, Will Deacon wrote:
> On Tue, May 14, 2024 at 11:00:50AM -0700, Namhyung Kim wrote:
>> for_each_sibling_event() checks leader's ctx but it doesn't have the ctx
>> yet if it's the leader. Like in perf_event_validate_size(), we should
>> skip checking siblings in that case.
>>
>> Fixes: f3c0eba287049 ("perf: Add a few assertions")
>> Reported-by: Greg Thelen <[email protected]>
>> Cc: Robin Murphy <[email protected]>
>> Cc: Tuan Phan <[email protected]>
>> Signed-off-by: Namhyung Kim <[email protected]>
>> ---
>> drivers/perf/arm_dmc620_pmu.c | 9 ++++++---
>> 1 file changed, 6 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/perf/arm_dmc620_pmu.c b/drivers/perf/arm_dmc620_pmu.c
>> index 8a81be2dd5ec..88c17c1d6d49 100644
>> --- a/drivers/perf/arm_dmc620_pmu.c
>> +++ b/drivers/perf/arm_dmc620_pmu.c
>> @@ -542,12 +542,16 @@ static int dmc620_pmu_event_init(struct perf_event *event)
>> if (event->cpu < 0)
>> return -EINVAL;
>>
>> + hwc->idx = -1;
>> +
>> + if (event->group_leader == event)
>> + return 0;
>> +
>> /*
>> * We can't atomically disable all HW counters so only one event allowed,
>> * although software events are acceptable.
>> */
>> - if (event->group_leader != event &&
>> - !is_software_event(event->group_leader))
>> + if (!is_software_event(event->group_leader))
>> return -EINVAL;

Oh, come to think of it, I believe we shouldn't actually need to keep
this check either, since commit bf480f938566 ("perf/core: Don't allow
grouping events from different hw pmus").

Thanks,
Robin.

>>
>> for_each_sibling_event(sibling, event->group_leader) {
>> @@ -556,7 +560,6 @@ static int dmc620_pmu_event_init(struct perf_event *event)
>> return -EINVAL;
>> }
>>
>> - hwc->idx = -1;
>> return 0;
>> }
>
> Thanks, I'll pick this up, although Mark reckoned he'd found some other
> issues over at:
>
> https://lore.kernel.org/r/Zg0l642PgQ7T3a8Z@FVFF77S0Q05N
>
> but didn't elaborate on what exactly he'd found :/
>
> Will
>
> _______________________________________________
> linux-arm-kernel mailing list
> [email protected]
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

2024-05-17 12:18:16

by Mark Rutland

[permalink] [raw]
Subject: Re: [PATCH] perf/arm-dmc620: Fix lockdep assert in ->event_init()

On Fri, May 17, 2024 at 01:02:34PM +0100, Will Deacon wrote:
> On Tue, May 14, 2024 at 11:00:50AM -0700, Namhyung Kim wrote:
> > for_each_sibling_event() checks leader's ctx but it doesn't have the ctx
> > yet if it's the leader. Like in perf_event_validate_size(), we should
> > skip checking siblings in that case.
> >
> > Fixes: f3c0eba287049 ("perf: Add a few assertions")
> > Reported-by: Greg Thelen <[email protected]>
> > Cc: Robin Murphy <[email protected]>
> > Cc: Tuan Phan <[email protected]>
> > Signed-off-by: Namhyung Kim <[email protected]>
> > ---
> > drivers/perf/arm_dmc620_pmu.c | 9 ++++++---
> > 1 file changed, 6 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/perf/arm_dmc620_pmu.c b/drivers/perf/arm_dmc620_pmu.c
> > index 8a81be2dd5ec..88c17c1d6d49 100644
> > --- a/drivers/perf/arm_dmc620_pmu.c
> > +++ b/drivers/perf/arm_dmc620_pmu.c
> > @@ -542,12 +542,16 @@ static int dmc620_pmu_event_init(struct perf_event *event)
> > if (event->cpu < 0)
> > return -EINVAL;
> >
> > + hwc->idx = -1;
> > +
> > + if (event->group_leader == event)
> > + return 0;
> > +
> > /*
> > * We can't atomically disable all HW counters so only one event allowed,
> > * although software events are acceptable.
> > */
> > - if (event->group_leader != event &&
> > - !is_software_event(event->group_leader))
> > + if (!is_software_event(event->group_leader))
> > return -EINVAL;
> >
> > for_each_sibling_event(sibling, event->group_leader) {
> > @@ -556,7 +560,6 @@ static int dmc620_pmu_event_init(struct perf_event *event)
> > return -EINVAL;
> > }
> >
> > - hwc->idx = -1;
> > return 0;
> > }
>
> Thanks, I'll pick this up, although Mark reckoned he'd found some other
> issues over at:
>
> https://lore.kernel.org/r/Zg0l642PgQ7T3a8Z@FVFF77S0Q05N
>
> but didn't elaborate on what exactly he'd found :/

Sorry; what I was referring to was that some drivers (including this
one) also forgot to validate that the group size could actually fit on
the PMU, which we're *supposed* to check, but if we fail to do so the
only fallout is that events won't count and we waste a bit of time
trying to schedule events unnecessarily.

For this patch as-is:

Acked-by: Mark Rutland <[email protected]>

.. and I'll try to take a look at the rest when I'm back in the UK.

Mark.

2024-05-17 16:18:07

by Will Deacon

[permalink] [raw]
Subject: Re: [PATCH] perf/arm-dmc620: Fix lockdep assert in ->event_init()

On Tue, 14 May 2024 11:00:50 -0700, Namhyung Kim wrote:
> for_each_sibling_event() checks leader's ctx but it doesn't have the ctx
> yet if it's the leader. Like in perf_event_validate_size(), we should
> skip checking siblings in that case.
>
>

Applied to arm64 (for-next/core), thanks!

[1/1] perf/arm-dmc620: Fix lockdep assert in ->event_init()
https://git.kernel.org/arm64/c/a4c5a457c610

Cheers,
--
Will

https://fixes.arm64.dev
https://next.arm64.dev
https://will.arm64.dev