[13441] in Athena Bugs
decmips 7.7M: nfs server
daemon@ATHENA.MIT.EDU (mhpower@MIT.EDU)
Tue Apr 25 03:41:01 1995
From: mhpower@MIT.EDU
To: bugs@MIT.EDU
Cc: sipb-staff@MIT.EDU
Date: Tue, 25 Apr 95 03:40:32 EDT
It appears possible to crash DECstations that serve nfs by writing a
large file into their nfs-exported directory from a Solaris machine.
On the DECstation (yaz-pistachio), I typed the following (this machine
isn't normally an nfs server, and I removed this nfs-server setup
after running the test that I'm describing here):
mkdir /var/test
echo /var/test > /etc/exports
/usr/athena/etc/mountd -i
/etc/nfsd 4
find /var -mount -exec /bin/ls -ld {} \; &
Then, on a Sun (running 7.7K), I typed
attach -n -m /mit/test1 -e yaz-pistachio:/var/test
cp testfile /mit/test1/testfile
where the size of testfile was 24 Mb. The file copy proceeded for
about 10 minutes, writing about 13 Mb into /mit/test1/testfile, and
then the DECstation crashed with:
EVENT CLASS ERROR EVENT
OS EVENT TYPE 200. PANIC
SEQUENCE NUMBER 19.
OPERATING SYSTEM ULTRIX 32
OCCURRED/LOGGED ON Tue Apr 25 02:10:20 1995 EDT
OCCURRED ON SYSTEM yaz-pistachi
SYSTEM ID x82070230 HW REV: x30
FW REV: x2
CPU TYPE: R2000A/R3000
PROCESSOR TYPE KN02-CA
PANIC MESSAGE smp_lock_long: beyond sleep count
This is a repeatable problem. It appears to affect all attempts to
write large files from Suns running Solaris over nfs to DECstations.
It may be necessary for the DECstation to have some other filesystem
activity (e.g., the find process in the above example) occurring at
the same time as the nfs copy, but I haven't confirmed that.
I've written files of similar length (over 20 Mb) to DECstation nfs
servers from other DECstations, and even from Suns running SunOS
4.1.3_U1, and this did not crash the DECstation nfs-server machine.
Writing large files from Solaris machines has crashed the DECstation
nfs-server machine in every case that I've tried (5 attempts, using
two different DECstations as servers). The panic message is always
"smp_lock_long: beyond sleep count". Prior to crashing, file-access
performance on the DECstation seems to degrade considerably. In
particular, trying to stat the file being created (e.g., by typing
"ls -l /var/test") is very slow (it takes about a minute to complete).
I'd be interested in whether the nfs implementation in later versions
of Ultrix (4.3 or 4.4) also exhibits this behavior.
Matt
