cvs commit: src/sys/dev/bce if_bce.c if_bcefw.h if_bcereg.h

[David Christensen]

> On 8.0/i386, with PAE enabled, I get messages on the console and the
> system hangs when trying to do a nfs mount.  Backing out the driver
> fixes it.  The same driver doesn't cause quite as spectacular a
> failure on 8.0/amd64, but it isn't exactly happy..
>
> Additional IP options:.^M
> Mounting NFS file systebcms:e1: link state changed to UP^M
> bce1: discard frame w/o leading ethernet header (len 0 pkt len 0)^M
> bce1: discard frame w/o leading ethernet header (len 0 pkt len 0)^M
> bce1: discard frame w/o leading ethernet header (len 0 pkt len 0)^M
> bce1: discard frame w/o leading ethernet header (len 0 pkt len 0)^M
> [..forever..]
>
> NFS over UDP, fwiw.  Server is a netapp.

I may have found the performance problem.  It looks like an issue
with handling the producer/consumer indices for the send ring that
would cause outgoing frames to just sit there until a received
frame came in and kickstarted the interrupt handler.  This is
consistent with your observation that enabling promiscuous mode
seemed to make things better, there's more received traffic.  I
found that pinging the SUT from another system has the same
effect.  Can you try the attached patch to head and let me know?

I tested on RELENG_7 and I haven't seen the discard errors yet.

Dave
-------------- next part --------------
A non-text attachment was scrubbed...
Name: if_bcereg.patch
Type: application/octet-stream
Size: 890 bytes
Desc: if_bcereg.patch
Url : http://lists.freebsd.org/pipermail/cvs-src/attachments/20080409/5ed4d565/if_bcereg.obj
-------------- next part --------------
A non-text attachment was scrubbed...
Name: if_bce.patch
Type: application/octet-stream
Size: 13993 bytes
Desc: if_bce.patch
Url : http://lists.freebsd.org/pipermail/cvs-src/attachments/20080409/5ed4d565/if_bce.obj