Re: [Intel-gfx] [PATCH v5 1/7] drm/i915/pmu: Change bitmask of enabled events to u32

["Dixit, Ashutosh" <ashutosh.dixit@intel.com>]

On Wed, 17 May 2023 01:26:15 -0700, Tvrtko Ursulin wrote:
>
>
> On 17/05/2023 07:55, Umesh Nerlige Ramappa wrote:
> > On Tue, May 16, 2023 at 05:25:50PM -0700, Dixit, Ashutosh wrote:
> >> On Tue, 16 May 2023 16:35:28 -0700, Umesh Nerlige Ramappa wrote:
> >>>
> >>
> >> Hi Umesh/Tvrtko,
> >>
> >> Mostly repeating comments/questions made on the previous patch below.
>
> First of all thanks for improving this, my v1 obviously wasn't good enough.
>
> >>
> >>> From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> >>>
> >>> Having it as u64 was a confusing (but harmless) mistake.
> >>>
> >>> Also add some asserts to make sure the internal field does not overflow
> >>> in the future.
> >>>
> >>> v2: Fix WARN_ON firing for INTERRUPT event (Umesh)
> >>>
> >>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
> >>> Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
> >>> Cc: Ashutosh Dixit <ashutosh.dixit@intel.com>
> >>> ---
> >>>  drivers/gpu/drm/i915/i915_pmu.c | 26 ++++++++++++++++++--------
> >>>  1 file changed, 18 insertions(+), 8 deletions(-)
> >>>
> >>> diff --git a/drivers/gpu/drm/i915/i915_pmu.c
> >>> b/drivers/gpu/drm/i915/i915_pmu.c
> >>> index 7ece883a7d95..96543dce2db1 100644
> >>> --- a/drivers/gpu/drm/i915/i915_pmu.c
> >>> +++ b/drivers/gpu/drm/i915/i915_pmu.c
> >>> @@ -50,7 +50,7 @@ static u8 engine_event_instance(struct perf_event
> >>> *event)
> >>>     return (event->attr.config >> I915_PMU_SAMPLE_BITS) & 0xff;
> >>>  }
> >>>
> >>> -static bool is_engine_config(u64 config)
> >>> +static bool is_engine_config(const u64 config)
> >>>  {
> >>>     return config < __I915_PMU_OTHER(0);
> >>>  }
> >>> @@ -88,9 +88,20 @@ static unsigned int config_bit(const u64 config)
> >>>         return other_bit(config);
> >>>  }
> >>>
> >>> -static u64 config_mask(u64 config)
> >>> +static u32 config_mask(const u64 config)
> >>>  {
> >>> -    return BIT_ULL(config_bit(config));
> >>> +    unsigned int bit = config_bit(config);
> >>
> >> Give that config_bit() can return -1 (I understand it is avoided in
> >> moving
> >> the code to config_mask from config_bit), maybe the code below should
> >> also
> >> have that check?
> >
> > config_mask is only called to check frequency related events in the code,
> > so I don't see it returing -1 here.
>
> Yeah that should be fine since -1 would make the below asserts fire
> anyway. (If it would get called from a different path in the future.)
>
> >>
> >>     int bit = config_bit(config);
> >>
> >>     if (bit != -1)
> >>     {
> >>         ...
> >>     }
> >>
> >> Though as mentioned below the 'if (__builtin_constant_p())' would have to
> >> go. Maybe the code could even have stayed in config_bit with the check.
> >>
> >>> +
> >>> +    if (__builtin_constant_p(config))
> >>> +        BUILD_BUG_ON(bit >
> >>> +                 BITS_PER_TYPE(typeof_member(struct i915_pmu,
> >>> +                             enable)) - 1);
> >>
> >> Given that config comes from the event (it is event->attr.config), can
> >> this
> >> ever be a builtin constant?
> >
> > Not sure about earlier code where these checks were inside config_bit(),
> > but with changes I made, I don't see this being a builtin
> > constant. However, nothing prevents a caller from just passing a
> > builtin_constant to this in future.
>
> Are you sure? I would have thought it would always be a compile time
> constant now that the check is in config_mask. Aahhh.. with the multi-tile
> changes maybe it can't unroll the loops and calculate the masks at compile
> time. Maybe it is a bit too much and we should drop the
> __builtin_constant_p branch? Probably..

Ah yes, with the code move to config_mask, they really all are compile time
constants (provided compiler can unroll the loops) so at least that is the
justfication for leaving the __builtin_constant_p in. So I'd probably just
leave it as is (though it is a bit too much).

> But I guess it is safe to use GEM_WARN_ON_ONCE instead of WARN_ON_ONCE
> since there are no external callers (nothing coming from event) now. That
> way at least production builds don't have to have the check.

Hmm, there's a GEM_WARN_ON but no GEM_WARN_ON_ONCE. So leave that as is too
I guess.

So I'm ok with the code staying as is. Enough bike-shed on this already.

Thanks.
--
Ashutosh


>
> Regards,
>
> Tvrtko
>
> >
> > Thanks,
> > Umesh
> >
> >>
> >>> +    else
> >>> +        WARN_ON_ONCE(bit >
> >>> +                 BITS_PER_TYPE(typeof_member(struct i915_pmu,
> >>> +                             enable)) - 1);
> >>
> >> There is really an even stricter limit on what the bit can be, which is
> >> the
> >> total number of possible events but anyway this is good enough.
> >>
> >> After addressing the above, this patch is:
> >>
> >> Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
> >>
> >>> +
> >>> +    return BIT(config_bit(config));
> >>>  }
> >>>
> >>>  static bool is_engine_event(struct perf_event *event)
> >>> @@ -633,11 +644,10 @@ static void i915_pmu_enable(struct perf_event
> >>> *event)
> >>>  {
> >>>     struct drm_i915_private *i915 =
> >>>         container_of(event->pmu, typeof(*i915), pmu.base);
> >>> +    const unsigned int bit = event_bit(event);
> >>>     struct i915_pmu *pmu = &i915->pmu;
> >>>     unsigned long flags;
> >>> -    unsigned int bit;
> >>>
> >>> -    bit = event_bit(event);
> >>>     if (bit == -1)
> >>>         goto update;
> >>>
> >>> @@ -651,7 +661,7 @@ static void i915_pmu_enable(struct perf_event
> >>> *event)
> >>>     GEM_BUG_ON(bit >= ARRAY_SIZE(pmu->enable_count));
> >>>     GEM_BUG_ON(pmu->enable_count[bit] == ~0);
> >>>
> >>> -    pmu->enable |= BIT_ULL(bit);
> >>> +    pmu->enable |= BIT(bit);
> >>>     pmu->enable_count[bit]++;
> >>>
> >>>     /*
> >>> @@ -698,7 +708,7 @@ static void i915_pmu_disable(struct perf_event
> >>> *event)
> >>>  {
> >>>     struct drm_i915_private *i915 =
> >>>         container_of(event->pmu, typeof(*i915), pmu.base);
> >>> -    unsigned int bit = event_bit(event);
> >>> +    const unsigned int bit = event_bit(event);
> >>>     struct i915_pmu *pmu = &i915->pmu;
> >>>     unsigned long flags;
> >>>
> >>> @@ -734,7 +744,7 @@ static void i915_pmu_disable(struct perf_event
> >>> *event)
> >>>      * bitmask when the last listener on an event goes away.
> >>>      */
> >>>     if (--pmu->enable_count[bit] == 0) {
> >>> -        pmu->enable &= ~BIT_ULL(bit);
> >>> +        pmu->enable &= ~BIT(bit);
> >>>         pmu->timer_enabled &= pmu_needs_timer(pmu, true);
> >>>     }
> >>>
> >>> --
> >>> 2.36.1
> >>>