Project

General

Profile

Actions

Bug #1695

open

NFS-related system breakdown

Added by Anonymous over 14 years ago. Updated over 10 years ago.

Status:
New
Priority:
Normal
Assignee:
-
Category:
-
Target version:
Start date:
Due date:
% Done:

0%

Estimated time:

Description

Hi all.

I do a cp(1) from /mnt/nfs (which is a zfs fileystem NFS-exported by an
opensolaris installation) to my /home/beket directory (HAMMER fs).

After a few MBs having being copied, cp(1) process stalls in 'getblk' state.
Also it is unkillable by ^C. From that point there's downhill.

Many commands will block if I issue them, such as 'mount' or 'df'. I can still
view top(1) update its contents but the system is on the edge. dmesg shows
sparse messages 'nfs server 10.0.0.1:/export/nfs: not responding' or
'[diagnostic]: $address block on cache_something ""'. If I break into the
debugger, I get nothing unusual. scgetc < sckbdevent < kbd_xxxx < taskqueue_yyyy
etc.

I tried to kill X and the system hang to the point that a cold reset was
necessary. I once managed to resume after typing 'c' in db> prompt, but that
doesn't always succeed.

This situation is, I think, 100% reproducible. Also, I don't have problems
copying stuff from|to a linux NFS client.

Cheers

Actions #1

Updated by Anonymous over 14 years ago

New data.

I hit ctrl+alt+delete and system stopped at 'unmount: there are still XXX
namecache references'. Since it stayed there for a few minutes, I hit again
ctrl+alt+delete and I got a kernel trap in nfsm_dissect. fault virtual address
is 0xdeadc0ea :)

Stack trace is nsfm_dissect < nfs_request < nsfsvc_iod_reader < lwkt_exit.

I think I got a kernel core dump. Will upload to leaf tomorrow.

Actions #2

Updated by Anonymous over 14 years ago

Dunno if this helps to bisect the offending code, but forcing mount_nfs to use
v2 protocol makes the problem go completely away. I have already copied the
video like 10 times without encountering a b/lock situation.

If I unmount the filesystem and remount it as v3 this time, the problem reappears.

Actions #3

Updated by dillon over 14 years ago

You could try a NFSv3 UDP mount instead of the default TCP mount to
see if that helps. If it doesn't then it should be possible to use
tcpdump to monitor the nfs traffic and figure out which rpc is stalling.

You can also mount with the 'intr' option which will make blocked
accesses interruptable.
There isn't enough information to determine who is at fault.
-Matt
Matthew Dillon
&lt;&gt;
Actions #4

Updated by Anonymous over 14 years ago

If I do NFSv3 UDP it works fine. I copied the file 5x with no b/lock failure. I
then remounted it as NFSv3 TCP and it failed as yesterday. I killed X and looked
at dmesg. I now saw some new messages:

EXDEV case 1 0xd1b06a88
EXDEV case 1 0xd1b06a88
EXDEV case 1 0xd1b06a88
EXDEV case 1 0xd1b06a88
[diagnostic] cache_lock: blocked on $addr ""
[diagnostic] cache_lock: blocked on $addr ""

Oh, and yes -i(ntr) did the trick. I was able to ^C the cp(1) and restore normal
operation.

What's next? tcpdump on the server side ?

Actions #5

Updated by dillon over 14 years ago

:
:What's next? tcpdump on the server side ?
:

I'm not sure tcpdump can track rpcs over a tcp connection, but that
would be the next step to try to find out which side is responsible
for dropping one of the RPCs on the floor.
Another possibility is that it is a stalled TCP connection which you
can detect by looking at the receive and transmit buffer backlog on
both sides via 'netstat -p tcp'.
-Matt
Matthew Dillon
&lt;&gt;
Actions #6

Updated by casusbubble over 10 years ago

  • Description updated (diff)

If I do NFSv3 UDP it works fine. I copied the file 5x with no b/lock failure.

http://www.reverse-your-diabetes-today.net/

Actions

Also available in: Atom PDF