Situation: Machine A with old ethernet card (3c503) produces "reciever ring
buffer overrun" some of the time due to (what I assumed to be) packets 
arriving before other packets were cleared from its buffer.  Access of 
files via NFS client or server would lock up if the files were larger than
 ~6K.

Reality: 8192K packets were being used by NFS and as they were too big for
machine A's buffer, they would be instantly dropped and caught in a loop
of resubmission/dropping.  "ring buffer" implies either a loaded net or
packets too large.

Solution: Mount filesystems with -is and -R5 so you can ^C out of locked
up NFS transfers, and limit the number of retries.

Use the -r and -w flags to limit the size of NFS packets (4096 works in my
case).

Define NFS_BOOT_RWSIZE=8192 when compiling a kernel, this will allow NFS to 
step down the transmit size when things aren't working.  It starts at that
size and steps down; so I could set NFS_BOOT_RWSIZE=4096 to make certain my
transfers would always work the first time (but I have other NFS mounts that
I'd like to run at 8K)

===

Is there a reason this last isn't defined by default?

-j