[lwip-devel] pbuf.c issues

[Christiaan Simons]

Hi,

* Can anyone explain the idea behind the
double locking in pbuf_pool_alloc() ??

I see both SYS_ARCH_PROTECT() macro's
and some pbuf.c local flags like pbuf_pool_alloc_lock
and pbuf_pool_free_lock.

I figured the pbuf.c local flags are only used
when SYS_LIGHTWEIGHT_PROT == 0,
but I can't see their purpose in a multi- or
single-threaded system.

For a multi threaded system it this might work somehow,
but for the single threaded case it only seems to slow
things down.

Doesn't it?

* I've done a little profiling with gprof and
it seems a lot of cycles are spent in the pbuf_alloc(),
pbuf_free() and pbuf_header() functions.

I've made some tiny changes to give some
compilation hints for a 'not-so-good' compiler,
but I guess the puzzle mentioned above may
have some influence as well.

There are some other tiny improvements that
could be made.  But I don't suspect this all adds
up to a significant improvement. I think most
of the pointer arithmatic may be the cause
of the relatively poor performance.

The user should disable LWIP_STATS
(thus PBUF_STATS) for production code.
I'll add this to the documentation.

* There seems to be an open issue
on memory alignment on some archs (MIPS?)
for this code.

This was reported for the 1.1.0 release,
but we didn't follow this up. A partial solution
was suggested. I suspect the MEM_ALIGN()
macro to increase the pointer arithmatic,
so I wont fix this for now.



Bye,

Christiaan Simons

Hardware Designer
Axon Digital Design

http://www.axon.tv