On Saturday 03 November 2012 15:22:26 David Miller wrote:
[...]
> Your packet layouts are very poorly designed and I want you to stop
> and think seriously about things before extending things further.

You are right about the packet design/layout. But I also have some problems 
with the following statement. Maybe you can help me to resolve it.

> All of this __packed stuff is a serious problem.
> 
> It means that on RISC system, fields such as your 32-bit sequence
> number, will be read and written using byte loads and stores.
> 
> This is terrible.
> 
> Instead, design the structures so that they are full filled out to
> at least 4 byte boundaries, so that they and the contents after
> them, are 4 byte aligned too.
> 
> Then you won't need to mark all of your packet header structs
> with __packed, and therefore the compiler can use full 32-bit
> loads and stores to access 32-bit fields.

Ok, lets assume batman-adv has everything 4 byte aligned and we are running on 
a machine without unaligned r/w access. The machine may issues bus errors and 
so on.

Now also assume following really unusual situation: We get our data from a 
ethernet driver and the skb stores the ethernet header. The start of the 
ethernet header is perfectly aligned (4 or even 16 byte boundary aligned). The 
the header is 14/18 byte long (6 byte src, 6 byte dst, 2 byte ethertype and 
maybe 4 byte vlan). Now the payload starts only on a 2 byte boundary -> it is 
never 4 byte boundary aligned. A 32 bit read now causes different variations 
of problems (reminder: bus error).

This brings me to the question: Why should the "32 bit align everything 
relative to the start of the struct" approach help to resolve the situation 
for the access of 32 bit data structure members? May I am missing some 
information?

Kind regards,
        Sven

Attachment: signature.asc
Description: This is a digitally signed message part.

Reply via email to