Re: [PATCH v2 1/9] phy: add phy_get_bus_width()/phy_set_bus_width()calls
From: Kishon Vijay Abraham I
Date: Mon Nov 04 2013 - 01:05:37 EST
Hi,
On Saturday 02 November 2013 11:28 PM, Tomasz Figa wrote:
On Saturday 02 of November 2013 13:47:09 Matt Porter wrote:
On Sat, Nov 02, 2013 at 10:46:55PM +0530, Kishon Vijay Abraham I wrote:
Hi Tomasz,
On Saturday 02 November 2013 06:44 PM, Tomasz Figa wrote:
Hi Matt,
On Friday 01 of November 2013 15:45:50 Matt Porter wrote:
This adds a pair of APIs that allows the generic PHY subsystem to
provide information on the PHY bus width. The PHY provider driver
may
use phy_set_bus_width() to set the bus width that the PHY supports.
The controller driver may then use phy_get_bus_width() to fetch the
PHY bus width in order to properly configure the controller.
I somehow does not like this. If we take this path for any further
properties that we may need, we will end up with a lot of consumer
specific properties stored in a PHY object having their own accessor
functions.
Only after all of us feel that a property is *generic* enough, we
allow it to be added in the PHY object.
I also want to note that this was discussed over in another thread [2]
where you did consider my rough stab at a more generic attribute
accessor. It was definitely my first reaction as the way to do it like
Tomasz has said. The specific accessors are more readable to me besides
the justification you mention above.
[2] http://lkml.indiana.edu/hypermail/linux/kernel/1310.3/00673.html
Personally I like that version much better, but still it would need to be
polished a bit.
How I imagine such interface to be implemented:
phy.h:
struct phy {
// ...
const struct phy_attrs *attrs;
// ...
};
static inline const struct phy_attrs *phy_get_attrs(struct phy *phy) {
return phy->attrs;
};
API's like get_attrs is loosely defined. I'd prefer to have fully
defined APIs. There might be other attributes which the consumer might
not be interested in. Instead of returning the entire structure, it
would be better if we have facilities for the consumer to request only
the required attributes.
phy driver:
static const struct phy_attrs my_phy_attrs = {
// ...
};
static int my_phy_probe(...)
{
// ...
phy = devm_phy_create_attrs(dev, &ops, &my_phy_attrs, NULL);
// ...
}
phy consumer:
// ...
const struct phy_attrs *phy_attrs;
phy_attrs = phy_get_attrs(phy);
// ...
Why I think it is better than what I've seen in this and previous instance
of this thread? (in random order)
a) Only the PHY driver can set the attrs.
b) PHY consumer has access only to a const pointer.
c) PHY attributes can be placed in a static struct inside a driver file,
without the need to call any functions to set particular attributes.
Agree with all your points for setting the attributes apart from the
fact that we won't be able to add any validation criteria for the
attributes while setting it if needed and also there won't be symmetric
APIs for getting and setting the attributes..
Cheers
Kishon
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/