On 1/28/2010 6:36 PM, Stephen Hemminger wrote:

Please try this patch (and only this patch), on 2.6.33-rc5[*];
none of the other patches that did not make it upstream because that
confuses things too much.

The code that checks for DMA mapping errors on receive buffers would
not handle errors correctly.  I doubt you have these errors, but if you
did then it would explain the problems.  The code has to be a little
tricky and build mapping for new rx buffer before releasing old one,
that way if new mapping fails, the old one can be reused.

If it works for you, I will resubmit with signed-off.

---
* If you want to use DMA debugging, then you will also need the match patch.
   

Ok - I'll also be running with the recent sched fork vs. hotplug vs. 
cpuset namespaces patch (commit 
fabf318e5e4bda0aca2b0d617b191884fda62703) from tip. Without that I get 
an rcu hang.

My plan then is to run with your patch, the rcu patch & the dma debug 
patch, but disable dma debug for now and see of the problem recurs. If 
it works, I'll know in a couple of days. If not, perhaps sooner :(.