On 3/25/08, Andy Fleming afleming@gmail.com wrote:
On Tue, Mar 25, 2008 at 11:33 AM, Stefan Roese sr@denx.de wrote:
On Tuesday 25 March 2008, Andy Fleming wrote:
I thought about this some more, and the problem is that
cpu_eth_init()
and board_eth_init() are mutually exclusive, with board_eth_init()
having
a higher priority. I think the following will work, but would
appreciate
some feedback.
I'm not sure that's necessarily the case. Imagine, for instance, an 85xx board that (for some reason) has on-board ethernet. I believe some of the DS systems do this. So the 85xx has 4 nics which the SOC knows how to initialize. But the board has an additional driver to init. Why not just allow them both?
Image a board that doesn't want all CPU (SoC) interfaces to get
initialized.
If for such a board a cpu-specific init routine exists, there is no
chance to
not enable (init) all those cpu interfaces as done in cpu_eth_init().
With this approach of mutually exclusive routines, it could define it's board_eth_init() and init only the Soc interfaces really needed. Plus additional ones of course.
Does this make sense?
Well, it makes sense to a degree. However, we already have a mechanism for enabling or disabling individual interfaces. The config file for each board can be used to determine which interfaces are setup by the cpu_eth_init() function. I don't really object to having them mutually exclusive. But I suspect the most common use case would become:
board_eth_init() { my_special_eth_init(bis);
cpu_eth_init(bis); }
And if everyone's going to do that, why bother making the functions mutually exclusive?
Andy
With the current incarnation of network code, that would be a common use case. However, I really don't like having interface-specific #defines in the config file, since it doesn't jibe well with kconfig and is just ugly. You know what I'm talking about - #define CONFIG_TSEC1_PHY_FLAGS etc. (I made this up, but I'm sure there are definitions like it) The main reason I went through this exercise is to move towards a point where all but the most trivial Ethernet controllers are initialized in board code, allowing us to pass PHY info, feature flags etc. on a per-interface basis.
regards, Ben