On Sat, 2023-01-14 at 01:06 +0800, Wang Yugui wrote:

Hi,

quoted

quoted

On Jan 13, 2023, at 09:41, Chuck Lever III
[off-list ref] wrote:

quoted

On Jan 12, 2023, at 4:30 AM, Wang Yugui
[off-list ref] wrote:

Hi,

quoted

Hi,

quoted

Hi,

We noticed a dead lock of 'umount.nfs4 /nfs/scratch -l'

reproducer:

mount /dev/sda1 /mnt/test/
mount /dev/sda2 /mnt/scratch/
systemctl restart nfs-server.service
mount.nfs4 127.0.0.1:/mnt/test/ /nfs/test/
mount.nfs4 127.0.0.1:/mnt/scratch/ /nfs/scratch/
systemctl stop nfs-server.service
umount -l /nfs/scratch #OK
umount -l /nfs/test #dead lock

Best Regards
Wang Yugui (wangyugui@e16-tech.com)
2023/01/11

quoted

kernel: 6.1.5-rc1

This problem happen on kernel 6.2.0-rc3+(upstream) too.

Can you clarify:

- By "deadlock" do you mean the system becomes unresponsive, or
that
 just the mount is stuck?

- Can you reproduce in a non-loopback scenario: a separate client
and
 server?

I’m not seeing how the use of the ‘-l’ flag is at all relevant
here. The exact same thing will happen if you don’t use ‘-l’. All
the latter does is hide the fact that it is happening from user
space.

As far as I’m concerned, this is pretty much expected behaviour
when you turn off the server before unmounting. It means that the
client can’t flush any remaining dirty data to the server and it
can’t clean up state. So just don’t do that?

In the case, 'df -h' will fail to work without the 'umount -l'.

so I thought we should make 'umount -l' to works.

The NFS filesystem doesn't know or care about the flags you use to call
the umount() system call. That's all handled by the VFS.
All NFS knows is that the VFS told it to clean up the super block
because it is no longer in use.

The calls to nfs4_proc_destroy_session() and nfs4_destroy_clientid()
will both eventually time out and allow the unmount to complete. So it
is not as if this is a permanent hang that forces you to reboot.

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com