Re: R: [lwip-users] ip_reass

[address@hidden]

Kieran Mansley wrote:
> On Fri, 2008-10-10 at 07:21 -0400, Ed Sutter wrote:
>   
> > Ok, I am running with the BF537, I'll take a look at that setting.
> > However, it still seems to me that this is a problem in the code
> > because the combination of...
> >
> >    1. MEM_ALIGNMENT = 4
> >    2. pbuf_alloc() setting payload buffers up to be aligned based on 
> > MEM_ALIGNMENT
> >    3. the increment of the payload pointer by 14 in ethernet_input()
> > and...
> >    4. the overlay of the "helper" structure in 
> > ip_reass_chain_frag_into_datagram_and_validate()
> >
> > will cause the misaligned access.  If the option you mention fixes it for my
> > situation on the BF537, that doesn't resolve it for other systems.
> > Right?
> >     
First of all, Ed, I think you're right on this, it seems like a bug to me.
> Including a 2-byte offset at the start of an ethernet frame in order to
> align the IP and TCP headers is pretty common, so it should work for
> other systems too.
It doesn't, in case the ethernet DMA engine needs aligned source (TX) or 
destination (RX) memory!
>   Documenting this need in the above circumstance
> would be handy.  As would making a check that the alignment requirements
> of the overlaid helper structure are met.  I'm surprised that we got
> that far though
That would be because when writing the code, I tested it only on an x86, 
which isn't very helpful in detecting these kind of errors :-(
>  - you would think that accesses to the IP header would
> have had the same problem in your case.
>   
The protocol headers are packed, which hides/solves this problem.
I'll add a bug entry. The possible fix will be packing the helper 
struct, I guess, since it has to fit inside the IP header - 
MEM_ALIGNMENT can make it too big.

Oh, Ed: thanks for finding this.

Simon