Re: [PATCH 2/7] perf: Create a symlink for a PMU

From: Greg KH <greg@kroah.com>
To: "Liang, Kan" <kan.liang@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>,
	peterz@infradead.org, mingo@redhat.com,
	linux-kernel@vger.kernel.org, eranian@google.com,
	namhyung@kernel.org, acme@kernel.org, jolsa@redhat.com
Subject: Re: [PATCH 2/7] perf: Create a symlink for a PMU
Date: Fri, 25 Jun 2021 07:17:54 +0200	[thread overview]
Message-ID: <YNVnAlCVc9BhBRZj@kroah.com> (raw)
In-Reply-To: <33bccec9-c4c4-d52b-9ee4-9a09e7353812@linux.intel.com>

On Thu, Jun 24, 2021 at 01:07:14PM -0400, Liang, Kan wrote:
> 
> 
> On 6/24/2021 11:31 AM, Greg KH wrote:
> > On Thu, Jun 24, 2021 at 08:24:29AM -0700, Andi Kleen wrote:
> > > 
> > > On 6/24/2021 7:29 AM, Greg KH wrote:
> > > > On Thu, Jun 24, 2021 at 07:24:31AM -0700, Andi Kleen wrote:
> > > > > > But first off, why is this symlink suddenly needed?  What is so special
> > > > > > about this new hardware that it breaks the existing model?
> > > > > The driver can be in two modes:
> > > > > 
> > > > > - Driver fully knows the hardware and puts in the correct Linux names
> > > > > 
> > > > > - Driver doesn't know the hardware but is in a fallback mode where it only
> > > > > looks at a discovery table. There we don't have the correct names, just an
> > > > > numeric identifier for the different hardware sub components.
> > > > Why does this matter?  Why would the driver not "know" the hardware?  If
> > > > it doesn't know it, why would it bind to it?
> > > 
> > > It's a similar concept as a PCI class. How to have a driver that can handle
> > > future hardware, but with some restrictions
> > 
> > But this is NOT how busses work in the driver model.
> > 
> > PCI classes are great, but we do NOT suddenly add a symlink in sysfs if
> > a driver goes from being handled by "generic_pci_type_foo" to
> > "vendor_foo".  Userspace can handle the change and life goes on.
> > 
> > > The perf CPU PMU has had a similar concept for a long time. The driver can
> > > be either in architectural mode (with a subset of features), or be fully
> > > enabled. This allows users who are on an older kernel to still use at least
> > > a subset of the functionality.
> > 
> > So a device name will move from "generic" to "specific", right?
> 
> We'd like to keep both names.

That is not how the driver model works, sorry.

Stick with the first name you got (i.e. the one in the kernel now).  Do
not try to use symlinks for things that should not be symlinked.

If a device changes names, wonderful, deal with that in userspace, we do
so all the time in tools for busses because device names are never
guaranteed to be the same each boot.  But bus names do not change for
the obvious reason that bus names are not dynamic, you pick them at
build time.

If you have userspace code that relies on device names to be static and
never change, then you need to stick with that as that was your choice
when you created that api.

thanks,

greg k-h