TCP/IP mbuf element size

[Eric Norum]

Charles Steaderman wrote:
> Hi all. I am working on an ethernet driver on an ARM CPU. When I am 
> requested to transmit data out of the ethernet port, I am given a chain 
> of mbufs. The chain is often times made up of many small buffers. For 
> example, when replying to a 1500 byte ping request, I receive 2 mbuf 
> chains. The first chain contains 3 elements of 14, 80, and 1420 bytes 
> respectfully. The second chain contains 3 elements of 34, 20, and 8 
> bytes respectfully. Unfortunately, the DMA controller on our chip does 
> not deal well with small buffers, and causes underflow errors in our 
> ethernet controller. Is there a way to tell the TCP/IP stack to give me 
> fewer, but larger, mbufs? I really don't want to have to copy the data 
> from the mbufs into a buffer prior to initiating a transfer.
> 

What's wrong with copying to a single mbuf before transmission?  Even if 
there were a way to tell the TCP/IP stack to give you a single mbuf, 
there TCP/IP stack would still have to copy from the packet fragment 
mbufs to the single mbuf so there would be no real improvement from 
having your driver do the copy itself.

-- 
Eric Norum                                 norume at aps.anl.gov
Advanced Photon Source                     Phone: (630) 252-2793
Argonne National Laboratory