NetBSD 5.0/-current: MBUF confusion ?

[Frank Kardel <kardel%netbsd.org@localhost>]

Hi,

I have observed following anomalies when copying large filesystems via 
rsync/ssh from one system (4.0 UP) to 5.0_BETA MP (Q9550 in i386 mode,
kernel with options MBUFTRACE to support pinpointing mbuf leakage,).

1) after boot already 1536 mbufs are allocated to 'unknown data'. This 
is a significant amount and leaves little room for real network buffers.
Downing unused interface shows that these 'unknown data' mbufs are 
presumably allocated for tx/rx rings, as the allocation is reduced when 
disabling an interface. This may not play well with many active 
interfaces and a limited / static wrt configuration number of mbufs.

2) Additionally I see after a while rsync's ssh quit with errors because of:
  - MAC corruption
  - bad packet length (with a ridiculous length value in the millions)
  This seems the more likely the less buffers are present.

Issue 1 should be checked for the relation of interfaces to mbuf 
configuration.

Issue 2 worries me the most - could it be that we have mbuf confusion
on low mbuf conditions that finds it way to user level causing data 
corruption ?

Frank