Re: [PATCH v8 1/3] perf: cavium: Support memory controller PMU counters

[Greg KH]

On Wed, Jul 26, 2017 at 02:02:42PM -0700, David Daney wrote:
> On 07/26/2017 01:08 PM, Greg KH wrote:
> > On Wed, Jul 26, 2017 at 01:02:38PM -0700, David Daney wrote:
> > > On 07/26/2017 10:33 AM, Greg KH wrote:
> > > > On Wed, Jul 26, 2017 at 06:30:49PM +0200, Borislav Petkov wrote:
> > > > > On Wed, Jul 26, 2017 at 09:19:49AM -0700, Greg KH wrote:
> > > > > > On Wed, Jul 26, 2017 at 05:55:48PM +0200, Borislav Petkov wrote:
> > > > > > > On Wed, Jul 26, 2017 at 05:45:15PM +0200, Jan Glauber wrote:
> > > > > > > > The PMU/EDAC devices are all PCI devices do I need the 'struct pci_dev *'.
> > > > > > > > I'm not aware of other ways to access these devices. Please enlighten
> > > > > > > > me if I'm missing something.
> > > > > > > 
> > > > > > > Me enlighten you on Cavium hardware?! You're funny.
> > > > > > > 
> > > > > > > So I don't know whether the PCI hotplug code can run more than one
> > > > > > > function upon PCI ID detection. Probably Greg will say, write a
> > > > > > > multiplexer wrapper. :-)
> > > > > > 
> > > > > > -ENOCONTEXT....
> > > > > > 
> > > > > > Anyway, pci questions are best asked on the linux-pci@vger list.  And
> > > > > > yes, all PCI devices end up with a 'struct pci_dev *' automatically.
> > > > > 
> > > > > Simple: so they have a PCI ID of a memory contoller and want to hotplug
> > > > > two drivers for it. And those two drivers should remain independent from
> > > > > each other.
> > > > 
> > > > Hahahahaha, no.  That's crazy, you were right in guessing what my answer
> > > > was going to be :)
> > > > 
> > > 
> > > 
> > > Just to be clear about the situation, the device is a memory controller.  It
> > > has two main behaviors we are interested in:
> > > 
> > > A) Error Detection And Correction (EDAC).  This should be connected to the
> > > kernel's EDAC subsystem.  An existing driver (drivers/edac/thunderx_edac.c)
> > > does exactly this.
> > > 
> > > B) Performance Counters for actions taken in the corresponding memory. This
> > > should be connected to the kernel's perf framework as an uncore-PMU (the
> > > subject of this patch set).
> > > 
> > > It is a single PCI device.  What should the driver architecture look like to
> > > connect it to two different kernel subsystems?
> > 
> > Modify the drivers/edac/thunderx_edac.c code to add support for
> > performance counters.
> > 
> 
> Thanks Greg.  This adds some clarity to the situation.
> 
> This technique does slightly complicate the mapping of files and directories
> in the kernel source tree to maintainers.
> 
> Also, if a given configuration disables CONFIG_EDAC there is some hackery
> needed to get the perf portion of the driver included.

Well, you all deserve it for trying to have a single PCI device do
multiple things at the same time.  There's no real good reason for
creating hardware that way, PCI devices are "free", you should go throw
a printed copy of the PCI spec at the firmware developers who did this
to you.

Then get them to fix the firmware so you have multiple PCI devices...

good luck!

greg k-h