[lwip-users] Re: [lwip] Packed structures

[Florian Schulze]

Hi!

I now only packed arp_hdr and ethip_hdr and it works. All bugs I had before also
disappeared. I will now make some tests on the possible speed, but it already
seems to be much faster than WatTCP32 on DOS. And I like lwIP much more, because
of the raw_api.

Thanks for the help and this great library!

Florian

----- Original Message -----
From: "Adam Dunkels" <address@hidden>
To: <address@hidden>
Sent: Wednesday, January 02, 2002 7:35 PM
Subject: Re: [lwip] Packed structures


> On Wednesday 02 January 2002 18.55, you wrote:
> > > I guess this depends on the compiler, but in my experience it is only the
> > > arp_hdr that needs to be packet. The other headers are aligned in such a
> > > way that 16-bit word accesses are on a 16-bit word boundary, 32-bit
> > > accesses are on a 32-bit boundary, etc. I guess the protocol designers
> > > had this in mind when they designed the header layouts.
> >
> > I will try that out, I hope it will not be aligned on 32bit boundaries.
>
> Incoming packets can be properly aligned by the device driver since the ARP
> header always is 14 bytes long. By leaving two bytes at the top of the pbuf,
> the IP header will always be 32-bit aligned. Since the IP header without IP
> options always is 20 bytes, the TCP header will be aligned as well and since
> the TCP header without options is 20 bytes, the TCP data will be aligned. (Of
> course, the TCP data might be misaligned anyway if the sender has segmented
> it in such a way.)
>
> Outgoing packets should always be aligned on 32-bit boundaries by setting the
> MEM_ALIGNMENT option to 4 in lwipopts.h and the PBUF_LINK_HLEN to 16 (the
> length of the aligned Ethernet header). Note that I haven't tested this in a
> system that actually traps misaligned memory references.
>
> > > When it comes to the best way to pack the structure, I would think that
> > > something along the lines of a PACK_STRUCT macro and a PACK_STRUCT_FIELD
> > > macro would be the way to go. The actual #definition of the macro could
> > > be put in the arch/cc.h file and the structs that needs packing could be
> > > defined as:
> > >
> > > PACK_STRUCT(
> > > struct a_struct_that_needs_to_be_packed {
> > >    PACK_STRUCT_FIELD(u8_t a);
> > >    PACK_STRUCT_FIELD(u32_t b);
> > > }
> > > )
> > >
> > > But it would looks rather cluttred...
> >
> > But it would only be for very few headers. The only problem I see with this
> > is that some compilers don't like #'s in #defines.
>
> You are right. Perhaps a scheme with two macros, PACK_STRUCT_PREAMBLE and
> PACK_STRUCT_POSTAMBLE or something like that, would be better.
>
> /adam
> --
> Adam Dunkels <address@hidden>
> http://www.sics.se/~adam
> [This message was sent through the lwip discussion list.]
>
>

[This message was sent through the lwip discussion list.]