On (20/08/26 18:16), Andy Shevchenko wrote:
> On Wed, Aug 26, 2020 at 11:49:20PM +0900, Sergey Senozhatsky wrote:
> > i2c, apparently, can match the same device twice - the first
> > time in ->match bus hook (i2c_device_match()), and the second
> > one in ->probe (i2c_device_probe()) bus hook.
> >
> > To make things more complicated, the second matching does not
> > do exactly same checks as the first one. Namely, i2c_device_match()
> > calls acpi_driver_match_device() which considers devices that
> > provide of_match_table and performs of_compatible() matching for
> > such devices. One important thing to note here is that ACPI
> > of_compatible() matching (acpi_of_match_device()) is part of ACPI
> > and does not depend on CONFIG_OF.
> >
> > i2c_device_probe(), on the other hand, calls acpi_match_device()
> > which does not perform of_compatible() matching, but instead
> > i2c_device_probe() relies on CONFIG_OF API to perform of_match_table
> > matching, IOW ->probe matching, unlike ->match matching, depends on
> > CONFIG_OF. This can break i2c device probing on !CONFIG_OF systems
> > if the device does not provide .id_table.
> >
> > i2c_device_probe()
> > ...
> > if (!driver->id_table &&
> > !i2c_acpi_match_device(dev->driver->acpi_match_table, client) &&
> > !i2c_of_match_device(dev->driver->of_match_table, client)) {
> > status = -ENODEV;
> > goto put_sync_adapter;
> > }
> >
> > i2c_of_match_device() on !CONFIG_OF systems is always false, so we never
> > perform of_match_table matching. i2c_acpi_match_device() does ACPI match
> > only, no of_compatible() matching takes place, even though the device
> > provides .of_match_table and ACPI is capable of matching such device.
> >
> > It is not entirely clear why the device is matched again in bus
> > ->probe after successful and proper matching in bus ->match. Let's
> > remove ->probe matching.
>
> Reviewed-by: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
> (assuming it's okay to go)
Thanks.
I tested the patch on x86_64 (a mix of i2c devices with and without
.id_table) and arm64 boards - didn't notice any difference, module
probing wise.
-ss
