Re: [PATCH v4 04/10] coresight: etm-perf: update to handle configuration selection

[Suzuki K Poulose]

On 3/4/21 2:19 PM, Mike Leach wrote:
> Hi Suzuki,
>
> On Thu, 4 Mar 2021 at 12:13, Suzuki K Poulose <suzuki.poulose@xxxxxxx> wrote:
> > On 1/28/21 5:09 PM, Mike Leach wrote:
> > > Loaded coresight configurations are registered in the cs_etm\cs_config sub
> > > directory. This extends the etm-perf code to handle these registrations,
> > > and the cs_syscfg driver to perform the registration on load.
> > >
> > > Signed-off-by: Mike Leach <mike.leach@xxxxxxxxxx>
> > > ---
> > >    .../hwtracing/coresight/coresight-config.h    |   5 +-
> > >    .../hwtracing/coresight/coresight-etm-perf.c  | 164 +++++++++++++++---
> > >    .../hwtracing/coresight/coresight-etm-perf.h  |   8 +
> > >    .../hwtracing/coresight/coresight-syscfg.c    |  13 +-
> > >    4 files changed, 166 insertions(+), 24 deletions(-)


> > > +static ssize_t etm_perf_cscfg_event_show(struct device *dev,
> > > +                                      struct device_attribute *dattr,
> > > +                                      char *buf)
> > > +{
> > > +     struct dev_ext_attribute *ea;
> > > +
> > > +     ea = container_of(dattr, struct dev_ext_attribute, attr);
> > > +     return scnprintf(buf, PAGE_SIZE, "%s\n", (const char *)(ea->var));
> > > +}
> >
> > "configid=0x%lx", (unsigned long)ea->var ?
>
> ea->var _is_ "configid=0x%lx" due to the way perf handles the events
> sub-dir entries.

This must be combined with the suggestion below.

> > > +
> > > +static int etm_perf_add_cscfg_event(struct device *dev, struct cscfg_config_desc *cs_cfg)
> > > +{
> > > +     struct dev_ext_attribute *ea;
> > > +     unsigned long hash;
> > > +     int ret;
> > > +     struct device *pmu_dev = etm_pmu.dev;
> > > +
> > > +     ea = devm_kzalloc(dev, sizeof(*ea), GFP_KERNEL);
> > > +     if (!ea)
> > > +             return -ENOMEM;
> > > +
> > > +     hash = (unsigned long)cs_cfg->id_ea->var;
> > > +
> > > +     sysfs_attr_init(&ea->attr.attr);
> > > +     ea->attr.attr.name = devm_kstrdup(dev, cs_cfg->name, GFP_KERNEL);
> > > +     if (!ea->attr.attr.name)
> > > +             return -ENOMEM;
> > > +
> > > +     /*
> > > +      * attribute value is "configid=<hash>".
> > > +      * this will be what perf evaluates when the config name is used
> > > +      * on the command line.
> > > +      */
> > > +     ea->var = devm_kzalloc(dev, CSCFG_EVENT_STR_SIZE, GFP_KERNEL);
> > > +     if (!ea->var)
> > > +             return -ENOMEM;
> >
> > Could we drop this string and use the "hash" instead ?
>
> No. My understanding is that we have added an events directory to
> cs_etm, and add the configurations in there:-
>
> cs_etm/events/autofdo
>
> Now the contents of autofdo are "configid=0x<hash-value>" - where
> hash-value is the hash of "autofdo".
>
> On the perf command line:-
>
> perf record -e cs_etm/autofdo/ .....
>
> will result in perf parsing autofdo, looking in the events dir for
> cs_etm, seeing the configid=-string, and parsing that to assign to
> configid attribute - which we have allocated to config2:63:32 - this
> will then appear as a value in the perf_event and we can load the
> configuration when starting up the event on the ETM etc.

Sorry, I was not explicit in my comments. You could drop the string and
have ea->var = hash. And the _show() could simply do

"configid=0x%lx" , hash

as mentioned above.

That would avoid another string allocation, with the same interface.

Suzuki