On Thu, Jul 31, 2008 at 09:44:17AM -0700, Chris Larson wrote:

quoted

quoted

was wondering if anyone had any advice on an issue a customer of ours 
is running into.  When they run a series of bandwidth tests, slowly 
increasing the bandwidth & # of tcp connections to roughly 30Mbit at 
4000 connections, their machine panics.  An address passed to 
dma_unmap_single is beyond lowmem, as though it was corrupted in some 
fashion.  The address is the data pointer in the ethernet dma 
descriptors, it pulls a descriptor out of the ring in the packet receipt 
functions and at that point the address is wrong.  Nowhere in the driver 
is a descriptor set up with that address, yet the datasheet on the 
mv643xx doesn't seem to indicate that the dma engine ever writes to the 
descriptors themselves, only to the data they point to (as you'd 
expect).  Does anyone have any ideas on this?  I'm afraid that I'm a bit 
out of my depth on this issue.  The architecture is ppc.
   

I've never seen this happen.  Which kernel version is this on?
 

One of MontaVista's 2.6.10 trees, too heavily modified to be of much use 
to you, I'm certain. Most of the recent mv643xx patches are in our tree, 
as one of our developers was the old maintainer.

Right.  Well, for what it's worth, I've never seen something like this
happening, but I've only tested ARM, and I only started looking at
mv643xx_eth in the ~ 2.6.25 timeframe, so that might not say much.

It doesn't occur in 2.4, not that that's very helpful :)

That would probably be an entirely different driver, wouldn't it? :)

I can try to reproduce with a more recent tree and let you know.
Thanks for your time.

If you can reproduce it, I would certainly be interested if you
could reproduce it on, say, 2.6.25 or later, since those versions
work for me.


thanks,
Lennert