On Wed, Jun 16, 2021 at 10:23:43AM -0500, Bjorn Helgaas wrote:
> On Wed, Jun 16, 2021 at 09:09:40AM +0800, liuqi (BA) wrote:
> > On 2021/6/12 7:33, Krzysztof Wilczyński wrote:
>
> > > > +static ssize_t hisi_pcie_event_sysfs_show(struct device *dev,
> > > > + struct device_attribute *attr, char *buf)
> > > > +{
> > > > + struct dev_ext_attribute *eattr;
> > > > +
> > > > + eattr = container_of(attr, struct dev_ext_attribute, attr);
> > > > +
> > > > + return sysfs_emit(buf, "config=0x%lx\n", (unsigned long)eattr->var);
> > > > +}
> > >
> > > I am not that familiar with the perf drivers, thus I might be completely
> > > wrong here, but usually for sysfs objects a single value is preferred,
> > > so that this "config=" technically would not be needed, unless this is
> > > somewhat essential to the consumers of this attribute to know what the
> > > value is? What do you think?
> >
> > "config=" is a supported for userspace tool, it is a kind of alias, so
> > cannot be remover here, thanks.
>
> I don't understand this part. This is brand-new functionality for the
> kernel, so there's no requirement to maintain compatibility for
> existing userspace tools.
>
> If there's a similar sysfs show function for other perf drivers, and
> you need to be compatible with *that*, fine. But if this is merely
> about being compatible with userspace that uses out-of-tree kernel
> functionality, that's not a real factor.
Right, I think this is standard for all perf drivers as it is how the perf
tool figures out how to select a given event in the perf_event_attr (which
has a 'config' field, which is what this refers to).
Will
