Re: [PATCH] perf/arm_smmuv3: Omit the two judgements which done in framework

[Robin Murphy]

On 19/02/2024 6:37 am, Yang Jialong 杨佳龙 wrote:
> 在 2024/2/10 0:32, Robin Murphy 写道:
> > On 2024-02-09 4:09 pm, Will Deacon wrote:
> > > On Thu, Dec 21, 2023 at 05:38:01PM +0800, JiaLong.Yang wrote:
> > > > 'event->attr.type != event->pmu->type' has been done in
> > > > core.c::perf_init_event() ,core.c::perf_event_modify_attr(), etc.
> > > >
> > > > This PMU is an uncore one. The core framework has disallowed
> > > > uncore-task events. So the judgement to event->cpu < 0 is no mean.
> > >
> > > It would be great to refer to the changes which added those checks to
> > > the perf core code. From reading the code myself, I can't convince 
> > > myself
> > > that perf_try_init_event() won't call into the driver.
> > >
> > > > The two judgements have been done in kernel/events/core.c
> > > >
> > > > Signed-off-by: JiaLong.Yang <jialong.yang@xxxxxxxxxxxx>
> > > > ---
> > > >   drivers/perf/arm_smmuv3_pmu.c | 8 --------
> > > >   1 file changed, 8 deletions(-)
> > >
> > > It looks like _many_ perf drivers have these checks, so if they really
> > > aren't needed, we can clean this up bveyond SMMU. However, as I said
> > > above, I'm not quite convinced we can drop them.
> >
> > Right, I think the logic prevents events with a specific PMU type 
> > being offered to other PMUs, but as far as I'm aware doesn't apply the 
> > other way round to stop generic events (PERF_TYPE_HARDWARE etc.) being 
> > offered to all PMUs, so it's those that system PMUs need to reject. >
> > It's been on my wishlist for a long time to have a capability flag to 
> > say "I don't handle generic events, please only ever give me events of 
> > my exact type" so we *can* truly factor this into the core.
>
>
> It's core framework's responsible to differ generic events and others, 
> or uncore pmu and core pmu.
> Here we have flag PERF_TYPE_HARDWARE, PERF_TYPE_HW_CACHE, 
> PERF_PMU_CAP_EXTENDED_HW_TYPE doing this.
>
> again:
>          rcu_read_lock();
>          pmu = idr_find(&pmu_idr, type);
>          rcu_read_unlock();
>          if (pmu) {
>                  if (event->attr.type != type && type != PERF_TYPE_RAW &&
>                      !(pmu->capabilities & PERF_PMU_CAP_EXTENDED_HW_TYPE))
>                          goto fail; /* generic event with no ability pmu */
> This can avoid driver code (event->attr.type != event->pmu->type).

Now consider the other case below that, where "type" has not matched 
anything registered, so "pmu" is NULL, and the event is then offered to 
*every* registered PMU to see if anyone accepts it. Note that even CPU 
PMUs don't always register as PERF_TYPE_RAW, and in particular arm_pmu 
doesn't.

Thanks,
Robin.