Arnd Bergmann wrote:
> Actually, we don't normally add the attribute((packed)) in cases like
> this one, where you already have manual padding in it. Marking this
> structure packed would only cause a small performance loss on accesses
> of its members on certain architectures, but not have an impact on
> correctness.
> 
> No architecture supported by Linux requires higher than natural alignment
> for any integer types, and a lot of other code would break otherwise.

That's not quite true.  Some architectures supported by Linux add
"external" padding to the size and alignment of a structure.

	struct my_subtype {
		unsigned char st1, st2;
	};

On Linux/ARM, sizeof(my_subtype) == 4 and __alignof__(my_subtype) == 4.
On Linux/x86, sizeof(my_subtype) == 2 and __alignof__(my_subtype) == 1.

This will break code which expects them to pack into an array, or
which is accessing this structure from a 2-byte aligned address.

This also effects structures containing other structures:

	struct my_type {
		unsigned char a, b, c, d;
		struct my_subtype st[2];
		unsigned char e, f, g, h;
	};

On Linux/ARM, sizeof(my_type) == 16 and __alignof__(my_subtype) == 4.
On Linux/ARM, sizeof(my_type) == 12 and __alignof__(my_subtype) == 1.

This did break one of my programs on Linux.  I had to decide between
using __attribute__((packed)) and losing portability, or stop using a
struct to access the data which was ugly but portable (the structures
had a lot more fields than this example).

-- Jamie