On Wed, 27 Jan 2010 10:34:51 -0500
Michael Breuer [off-list ref] wrote:

On 01/23/2010 06:21 PM, Jarek Poplawski wrote:

quoted

On Fri, Jan 22, 2010 at 06:50:21PM -0500, Michael Breuer wrote:
     

quoted

When the packets were dropped, there was a different sequence in the
log - DISCOVER/OFFER repeated. The "normal" is that the sequence
appeared correct and complete - DISCOVER/OFFER/REQUEST/ACK - or
INFORM/ACK (vs. INFORM repeatedly sans ACK) as the case may be.
       

Anyway, I'd be intersted if the switch matters here.

Plus one more test: could you try to load sky2 with the parameter:
"copybreak=1" (the rest as in any recent test, which gave you dmar
errors; any switch).

Thanks,
Jarek P.
     

Ok - now up 80+ hours with copybreak=1. I'm going to redo w/o copybreak 
to confirm that I haven't inadvertently fixed something. However, given 
that it might be copybreak-related, I looked at sky2.c again and I'm 
wondering about the copybreak max size in sky2_rx_start:

   size = roundup(sky2->netdev->mtu + ETH_HLEN + VLAN_HLEN, 8);

         /* Stopping point for hardware truncation */
         thresh = (size - 8) / sizeof(u32);

         sky2->rx_nfrags = size >> PAGE_SHIFT;
         BUG_ON(sky2->rx_nfrags > ARRAY_SIZE(re->frag_addr));

         /* Compute residue after pages */
         size -= sky2->rx_nfrags << PAGE_SHIFT;

         /* Optimize to handle small packets and headers */
         if (size < copybreak)
                 size = copybreak;
         if (size < ETH_HLEN)
                 size = ETH_HLEN;


Why would increasing size to copybreak be valid here?

Guessing a bit as I'm not sure about rx_nfrags, but if I read this 
correctly, if size is ever less than copybreak it's because there isn't 
enough space left for anything larger. If so, wouldn't increasing size 
potentially corrupt something? I'd further guess that the resulting 
condition manifests sooner (or at least with a more visible effect) when 
using DMAR.

In any event, why "copybreak" as the minimum buffer size? I'd suggest 
that if it isn't possible to allocate at least MTU + overhead that 
sky2_rx_start ought to be delayed until there is room.

This code is where driver decides how much data will be received in skb
data area and the remaining data spills over into skb frags.
Copybreak is the threshold so that packets less than size are copied
to a new skb.  The code doing the copying there assumes the data is
totally contained in the skb (not in frags). The size increase there
is to make sure that assumption is always true.  I suppose you
could do something perverse like setting copybreak really huge 
and confuse driver, but that is a user error.