On Sat, Jul 11, 2020 at 06:46:01AM -0700, Richard Cochran wrote:
> On Fri, Jul 10, 2020 at 01:58:44PM -0700, Jakub Kicinski wrote:
> > On Fri, 10 Jul 2020 11:41:25 -0400 min.li...@renesas.com wrote:
> > > From: Min Li <min.li...@renesas.com>
> > >
> > > This patch is to add debugfs support for ptp_clockmatrix and ptp_idt82p33.
> > > It will create a debugfs directory called idtptp{x} and x is the ptp
> > > index.
> > > Three inerfaces are present, which are cmd, help and regs. help is read
> > > only and will display a brief help message. regs is read only amd will
> > > show
> > > all register values. cmd is write only and will accept certain commands.
> > > Currently, the accepted commands are combomode to set comobo mode and
> > > write
> > > to write up to 4 registers.
> > >
> > > Signed-off-by: Min Li <min.li...@renesas.com>
> >
> > No private configuration interfaces in debugfs, please.
>
> I suggested to Min to use debugfs for device-specific configuration
> that would be out of place in the generic PTP Hardware Clock
> interface.
>
> > If what you're exposing is a useful feature it deserves a proper
> > uAPI interface.
>
> Can you expand on what you mean by "proper uAPI interface" please?
Hi Richard
Well, one obvious issues is that debugfs it totally optional, and
often not built. You would not want the correct operation of a device
to depend something which is optional. debugfs is also unstable. There
are no ABI rules. So user space cannot rely on the API being the same
from version to version. Again, not something you want for the correct
operation of a device.
Allowing registers to be read, is a typical debug operation. So that
part seems reasonable. A kernel developer/debugger has the skills to
deal with unstable APIs, and rebuilding the kernel to actually have
debugfs in the image.
But configuration does not belong in debugfs. It would be good to
explain what is being configured by these parameters, then we can
maybe make a suggestion about the correct API to use.
Andrew
