Re: [EXT] Re: [PATCH v8 3/8] perf: imx_perf: let the driver manage the counter usage rather the user

[Frank Li <Frank.li@nxp.com>]

On Wed, Apr 10, 2024 at 04:45:47PM +0100, Will Deacon wrote:
> On Wed, Apr 10, 2024 at 07:39:46AM +0000, Xu Yang wrote:
> > > On Fri, Mar 22, 2024 at 02:39:25PM +0800, Xu Yang wrote:
> > > > In current design, the user of perf app needs to input counter ID to count
> > > > events. However, this is not user-friendly since the user needs to lookup
> > > > the map table to find the counter. Instead of letting the user to input
> > > > the counter, let this driver to manage the counters in this patch.
> > > 
> > > I think we still have to support the old interface so that we don't break
> > > those existing users (even if the driver just ignores whatever counter ID
> > > is provided in a backwards-compatible way).
> > > 
> > > > This will be implemented by:
> > > >  1. allocate counter 0 for cycle event.
> > > >  2. find unused counter from 1-10 for reference events.
> > > >  3. allocate specific counter for counter-specific events.
> > > >
> > > > In this patch, counter attribute is removed too. To mark counter-specific
> > > > events, counter ID will be encoded into perf_pmu_events_attr.id.
> > > >
> > > > Reviewed-by: Frank Li <Frank.Li@nxp.com>
> > > > Signed-off-by: Xu Yang <xu.yang_2@nxp.com>
> > > >
> > > > ---
> > > > Changes in v6:
> > > >  - new patch
> > > > Changes in v7:
> > > >  - no changes
> > > > Changes in v8:
> > > >  - add Rb tag
> > > > ---
> > > >  drivers/perf/fsl_imx9_ddr_perf.c | 168 ++++++++++++++++++-------------
> > > >  1 file changed, 99 insertions(+), 69 deletions(-)
> > > >
> > > > diff --git a/drivers/perf/fsl_imx9_ddr_perf.c b/drivers/perf/fsl_imx9_ddr_perf.c
> > > > index 0017f2c9ef91..b728719b494c 100644
> > > > --- a/drivers/perf/fsl_imx9_ddr_perf.c
> > > > +++ b/drivers/perf/fsl_imx9_ddr_perf.c
> > > > @@ -245,14 +249,12 @@ static const struct attribute_group ddr_perf_events_attr_group = {
> > > >       .attrs = ddr_perf_events_attrs,
> > > >  };
> > > >
> > > > -PMU_FORMAT_ATTR(event, "config:0-7");
> > > > -PMU_FORMAT_ATTR(counter, "config:8-15");
> > > > +PMU_FORMAT_ATTR(event, "config:0-15");
> > > 
> > > Sadly, this is a user-visible change so I think it will break old tools,
> > > won't it?
> > 
> > For imx ddr pmu, most of the people will use metrics rather event itself,
> > and we have speficy the format of event parameters in metrics table.
> > The parameters is also updated in this patchset.
> > 
> > And to easy use for user, the counter should be hidden (transparent) to
> > user after I had talk with Frank. Then, the user need't to look up the event
> > table to find which counter to use. 
> > 
> > So this patchset will basically not break the usage of perf tools and will
> > improve practicality.
> 
> Sorry, but I don't agree. The original commit adding this driver
> (55691f99d417) gives the following examples in the commit message:
> 
> For example:
>       perf stat -a -I 1000 -e imx9_ddr0/eddrtq_pm_rd_trans_filt,counter=2,axi_mask=ID_MASK,axi_id=ID/
>       perf stat -a -I 1000 -e imx9_ddr0/eddrtq_pm_wr_trans_filt,counter=3,axi_mask=ID_MASK,axi_id=ID/
>       perf stat -a -I 1000 -e imx9_ddr0/eddrtq_pm_rd_beat_filt,counter=4,axi_mask=ID_MASK,axi_id=ID/
> 
> I don't think these will work any more if we apply this patch.

Yang:

    keep compatible is important. Please try well's suggestion
    "if the driver just ignores whatever counter ID is provided in a 
backwards-compatible way)." 

    If you have further question, you can ping directly. Please avoid
use outlook to reply community email. It always append annoised "EXT".

Frank

> 
> Will