I've found what seems to be a reliable way to corrupt the buffer cache
for NFS filesystems in recent kernels. (Verified in 1.3.81 and 1.3.82;
i think it was present in 1.3.76 as well.)
I start up a large binary from the NFS filesystem for the first time.
Before the program starts up, i interrupt it with control-C.
>From then on, trying to run the binary again gives a segmentation fault.
(After a reboot, the binary runs fine as long as i don't interrupt it
the first time.)
After going through this procedure, i tried copying the binary to a local
disk, rebooting, and then comparing the copy with the original. They were,
in fact, different.
I won't have a chance to look into this more deeply for at least a few
days, but maybe this hint will help someone else experiencing the
same problem...
sss
snyder@fnald0.fnal.gov