Re: [PATCH v4 2/2] perf x86: Exposing an Uncore unit to PMON for Intel Xeon® server platform

[Greg KH]

On Tue, Jan 21, 2020 at 07:15:56PM +0300, Sudarikov, Roman wrote:
> On 17.01.2020 19:54, Greg KH wrote:
> > On Fri, Jan 17, 2020 at 08:23:57AM -0800, Andi Kleen wrote:
> > > > I thought I was nice and gentle last time and said that this was a
> > > > really bad idea and you would fix it up.  That didn't happen, so I am
> > > > being explicit here, THIS IS NOT AN ACCEPTABLE FILE OUTPUT FOR A SYSFS
> > > > FILE.
> > > Could you suggest how such a 1:N mapping should be expressed instead in
> > > sysfs?
> > I have yet to figure out what it is you all are trying to express here
> > given a lack of Documentation/ABI/ file :)
> > 
> > But again, sysfs is ONE VALUE PER FILE.  You have a list of items here,
> > that is bounded only by the number of devices in the system at the
> > moment.  That number will go up in time, as we all know.  So this is
> > just not going to work at all as-is.
> > 
> > greg k-h
> 
> Hi Greg,
> 
> Technically, the motivation behind this patch is to enable Linux perf tool
> to attribute IO traffic to IO device.
> 
> Currently, perf tool provides interface to configure IO PMUs only without
> any
> context.
> 
> Understanding IIO stack concept to find which IIO stack that particular
> IO device is connected to, or to identify an IIO PMON block to program
> for monitoring specific IIO stack assumes a lot of implicit knowledge
> about given Intel server platform architecture.

Is "IIO" being used here the same way that drivers/iio/ is in the
kernel, or is this some other term?  If it is the same, why isn't the
iio developers involved in this?  If it is some other term, please
always define it and perhaps pick a different name :)

> Please consider the following mapping schema:
> 
> 1. new "mapping" directory is to be added under each uncore_iio_N directory

What is uncore_iio_N?  A struct device?  Or something else?

> 2. that "mapping" directory is supposed to contain symlinks named "dieN"
> which are pointed to corresponding root bus.
> Below is how it looks like for 2S machine:
> 
> # ll uncore_iio_0/mapping/
> lrwxrwxrwx 1 root root 0 Jan 20 23:55 die0 ->
> ../../pci0000:00/pci_bus/0000:00

Where did "pci_bus" come from in there?  I don't see under /sys/devices/
for my pci bridges.

> lrwxrwxrwx 1 root root 0 Jan 20 23:55 die1 ->
> ../../pci0000:80/pci_bus/0000:80
> 
> # ll uncore_iio_1/mapping/
> lrwxrwxrwx 1 root root 0 Jan 20 23:55 die0 ->
> ../../pci0000:17/pci_bus/0000:17
> lrwxrwxrwx 1 root root 0 Jan 20 23:55 die1 ->
> ../../pci0000:85/pci_bus/0000:85
> 
> # ll uncore_iio_2/mapping/
> lrwxrwxrwx 1 root root 0 Jan 20 23:55 die0 ->
> ../../pci0000:3a/pci_bus/0000:3a
> lrwxrwxrwx 1 root root 0 Jan 20 23:55 die1 ->
> ../../pci0000:ae/pci_bus/0000:ae
> 
> # ll uncore_iio_3/mapping/
> lrwxrwxrwx 1 root root 0 Jan 20 23:55 die0 ->
> ../../pci0000:5d/pci_bus/0000:5d
> lrwxrwxrwx 1 root root 0 Jan 20 23:55 die1 ->
> ../../pci0000:d7/pci_bus/0000:d7

Why have a subdir here?

Anyway, yes, that would make sense, if userspace can actually do
something with that, can it?

Also, what tears those symlinks down when you remove those pci devices
from the system?  Shouldn't you have an entry in the pci device itself
for this type of thing?  And if so, isn't this really just a "normal"
class type driver you are writing?  That should handle all of the
symlinks and stuff for you automatically, right?

thanks,

greg k-h