On Mon, Oct 23, 2006 at 07:59:40PM -0500, John Darrow wrote:
> On 20 Oct 2006 17:21:45 -0500, Peter Eisch <peter@boku.net> wrote:
> >
> >Oct 20 15:15:14: vcpu-0| NOT_IMPLEMENTED
> >/build/mts/release/bora-29996/pompeii2005/bora/devices/net/vlance_shared.c:6
> >98
> 
> Yes, we have encountered this.  According to gkm (who works/worked at
> VMWare), it's a VMWare bug with the vlance emulation only supporting
> "15 sg elements in esx".  We were also seeing hangs of output from
> individual NetBSD VMs (e.g. a ssh session would hang when hit with
> lots of output).
> 
> The included patch seems to be handling things for us (we haven't
> seen crashes or hangs since it was put in place).
> 
> --- /current/src/sys/dev/pci/if_pcn.c_v1.32	2006-10-16 19:28:36.000000000 -0500
> +++ /current/src/sys/dev/pci/if_pcn.c	2006-10-18 02:17:04.000000000 -0500
> @@ -126,7 +126,11 @@
>   * DMA segments, but only allocate the max of 512 descriptors.  The
>   * transmit logic can deal with this, we just are hoping to sneak by.
>   */
> +#ifdef PCN_VMWARE_SGLIMIT
> +#define	PCN_NTXSEGS		4

I think you want "15" here, not "4".  More is better, and according to
the information above, 15 should be safe, no?

It's bogus for the thing to claim to be a pcnet if it in fact is not
quite a pcnet, and since the number of supported s/g elements is a
documented in the pcnet and ilacc databooks... sigh.

If you don't want to patch your kernel, it might be possible to run the
thing in 24-bit-descriptor mode instead by removing pcn from your kernel
so it matches le@pci.  I wonder if that will show the same Vmware bug?

Thor