Re: [PATCH v2 0/6] PCI: Drop duplicated tracking of a pci_dev's bound driver

[Bjorn Helgaas <helgaas@kernel.org>]

On Fri, Aug 06, 2021 at 08:46:23AM +0200, Uwe Kleine-König wrote:
> On Thu, Aug 05, 2021 at 06:42:34PM -0500, Bjorn Helgaas wrote:

> > I looked at all the bus_type.probe() methods, it looks like pci_dev is
> > not the only offender here.  At least the following also have a driver
> > pointer in the device struct:
> > 
> >   parisc_device.driver
> >   acpi_device.driver
> >   dio_dev.driver
> >   hid_device.driver
> >   pci_dev.driver
> >   pnp_dev.driver
> >   rio_dev.driver
> >   zorro_dev.driver
> 
> Right, when I converted zorro_dev it was pointed out that the code was
> copied from pci and the latter has the same construct. :-)
> See
> https://lore.kernel.org/r/20210730191035.1455248-5-u.kleine-koenig@pengutronix.de
> for the patch, I don't find where pci was pointed out, maybe it was on
> irc only.

Oh, thanks!  I looked to see if you'd done something similar
elsewhere, but I missed this one.

> > Looking through the places that care about pci_dev.driver (the ones
> > updated by patch 5/6), many of them are ... a little dubious to begin
> > with.  A few need the "struct pci_error_handlers *err_handler"
> > pointer, so that's probably legitimate.  But many just need a name,
> > and should probably be using dev_driver_string() instead.
> 
> Yeah, I considered adding a function to get the driver name from a
> pci_dev and a function to get the error handlers. Maybe it's an idea to
> introduce these two and then use to_pci_driver(pdev->dev.driver) for the
> few remaining users? Maybe doing that on top of my current series makes
> sense to have a clean switch from pdev->driver to pdev->dev.driver?!

I'd propose using dev_driver_string() for these places:

  eeh_driver_name() (could change callers to use dev_driver_string())
  bcma_host_pci_probe()
  qm_alloc_uacce()
  hns3_get_drvinfo()
  prestera_pci_probe()
  mlxsw_pci_probe()
  nfp_get_drvinfo()
  ssb_pcihost_probe()

The use in mpt_device_driver_register() looks unnecessary: it's only
to get a struct pci_device_id *, which is passed to ->probe()
functions that don't need it.

The use in adf_enable_aer() looks wrong: it sets the err_handler
pointer in one of the adf_driver structs.  I think those structs
should be basically immutable, and the drivers that call
adf_enable_aer() from their .probe() methods should set
".err_handler = &adf_err_handler" in their static adf_driver
definitions instead.

I think that basically leaves these:

  uncore_pci_probe()     # .id_table, custom driver "registration"
  match_id()             # .id_table, arch/x86/kernel/probe_roms.c
  xhci_pci_quirks()      # .id_table
  pci_error_handlers()   # roll-your-own AER handling, drivers/misc/cxl/guest.c

I think it would be fine to use to_pci_driver(pdev->dev.driver) for
these few.

Bjorn