Re: [External] Re: [PATCH] perf/dwc_pcie: fix duplicate pci_dev devices

[Will Deacon]

On Sat, Feb 01, 2025 at 05:51:21PM +0800, yunhui cui wrote:
> On Tue, Jan 28, 2025 at 12:52 AM Will Deacon <will@xxxxxxxxxx> wrote:
> > On Sun, Jan 26, 2025 at 09:54:35AM +0800, yunhui cui wrote:
> > > On Fri, Jan 24, 2025 at 8:30 PM Will Deacon <will@xxxxxxxxxx> wrote:
> > > >
> > > > On Fri, Jan 24, 2025 at 05:11:05PM +0800, yunhui cui wrote:
> > > > > > >>>        vsec = dwc_pcie_des_cap(pdev);
> > > > > > >>>        if (!vsec)
> > > > > > >>>                return -ENODEV;
> > > > > > >>>
> > > > > > >>>        sbdf = plat_dev->id;
> > > > > > >>> -     name = devm_kasprintf(&plat_dev->dev, GFP_KERNEL, "dwc_rootport_%x", sbdf);
> > > > > > >>> +     name = devm_kasprintf(&plat_dev->dev, GFP_KERNEL, "dwc_rootport_%d_pmu", sbdf);
> > > > > > >>
> > > > > > >> A new name will break previous user tools.
> > > > > > > This name isn't suitable. It can't clearly show which is the PMU
> > > > > > > device. Userspace tools don't have binding relationships, like perf.
> > > > > > > Tools must traverse PMU devices before use.
> > > > > >
> > > > > > The device is under /sys/bus/event_source/ which indates it is PMU device.
> > > > > > As far as I know, most of PMU devices do not endup with a '_pmu' prefix.
> > > > >
> > > > > The point is the name "dwc_rootport_" is misleading, suggesting an
> > > > > extra PCIE RP in the system. Best to change the name to intuitive
> > > > > "xx_pmu".
> > > >
> > > > As pointed out above, changing the name will break userspace. So it's
> > > > simply not an option, sorry.
> > >
> > > Could you specify what in the userspace is broken and give an example?
> >
> > Can't I just take the documented example from dwc_pcie_pmu.rst:
> >
> > $# perf stat -a -e dwc_rootport_13018/Rx_PCIe_TLP_Data_Payload/
> >
> > If I have a script that uses that command on a machine today, then
> > applying your kernel patch will break the script, no?
> 
> Well, this is just a usage guide for the PCIe PMU and has no binding
> with perf. In user space, the existing PCIe PMUs in the system should
> be traversed first, and then perf stat/record ... can be executed.

"should be". ha! You know we can't rely on userspace doing whatever we
think it "should" do.

> It's okay not to change the name, but it's better to change
> dwc_rootport_%x to dwc_rootport_%d to correspond with
> platform_device_register_data(NULL, "dwc_pcie_pmu", ...). It'll be
> more intuitive. What do you think?

No, as I said earlier, we cannot break userspace like this. Please stop
trying to rename things that cannot be changed.

Will