Re: drivers/net/wireless/rt2x00/: struct data_desc strangeness

[Ivo van Doorn]

On Sunday 14 October 2007, Adrian Bunk wrote:
> drivers/net/wireless/rt2x00/rt2x00ring.h contains the following:
> 
> <--  snip  -->
> 
> ...
> /*
>  * data_desc
>  * Each data entry also contains a descriptor which is used by the
>  * device to determine what should be done with the packet and
>  * what the current status is.
>  * This structure is greatly simplified, but the descriptors
>  * are basically a list of little endian 32 bit values.
>  * Make the array by default 1 word big, this will allow us
>  * to use sizeof() correctly.
>  */
> struct data_desc {
>         __le32 word[1];
> };
> ...
> /*
>  * TX/RX Descriptor access functions.
>  */
> static inline void rt2x00_desc_read(struct data_desc *desc,
>                                     const u8 word, u32 *value)
> {
>         *value = le32_to_cpu(desc->word[word]);
> }
> 
> static inline void rt2x00_desc_write(struct data_desc *desc,
>                                      const u8 word, const u32 value)
> {
>         desc->word[word] = cpu_to_le32(value);
> }
> ...
> 
> <--  snip  -->
> 
> I haven't checked whether it might work in all cases, but passing 
> non-zero values as second parameter to rt2x00_desc_{read,write}()
> (as is done in many cases) is even in the best case bad coding style.

Access to the array is correct, even with non-zero values the code that is
reading/writing to the array knows the exact size of the descriptor. Within
rt2x00 are however 5 drivers who have different descriptor sizes. That means
I can't create a structure which has the correct array length.

The structure itself is just a simple map over preallocated memory
(skb->data in case of USB or dma in case of PCI).

So possible solutions would be:
 - remove struct data_desc and make it a void* or __le32*
   This is at the cost of code readibility since the above pointers
   have less meaning then a pointer to a structure which can be nicely
   documented.
 - increase the word[] array to something that fits all (+/- 20 entries)
   This wouldn't really be a problem, all it requires is fixing the sizeof() 
   statements. But then the code should contain a big note about that it
   is not allowed to read/write _all_ words in the entry since it depends on
   the driver that uses it.

What would in this case be the best solution for good coding style?

Ivo
-
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/