NFS bug with child datasets

[gizimoto]

After seeing some odd behaviour with my NFS shares (http://forums.freenas.org/showthread.php?10655) I believe I have found a bug.

Child datasets under NFS sharing do not correctly write data to the ZFS dataset, rather to a normal directory.

Here is how to repeat the issue:

Create ZFS data set with child dataset, for example
/mnt/storage/test
/mnt/storage/test/testchild

Create an NFS share on the higher level /mnt/storage/test location.

Mount the NFS share, I use Ubuntu and AutoFS, but 'mount -t nfs 192.168.1.205:/mnt/storage/test /mnt/test' shows same behaviour.

Any files/folders created in the test/testchild folder do not appear when viewing the files directly from the FreeNAS shell (nor from a CIFS share which is how I found the issue).

For example from my Ubuntu server:

Code:

mark@ubuntuserver:/mnt/test$ mount -l
... snip ...
192.168.1.205:/mnt/storage/test on /mnt/test type nfs (rw,addr=192.168.1.205)
mark@ubuntuserver:/mnt/test$ ls -la
total 38
drwxrwxr-x 3 root 1001 3 Jan  6 11:08 .
drwxr-xr-x 5 root root 0 Jan  6 11:03 ..
drwxr-xr-x 2 root 1001 3 Jan  6 11:09 testchild
mark@ubuntuserver:/mnt/test$ cd testchild/
mark@ubuntuserver:/mnt/test/testchild$ vi from_ubuntuserver_nfs.txt
mark@ubuntuserver:/mnt/test/testchild$ ls -la
total 48
drwxr-xr-x 2 root 1001 3 Jan  6 11:09 .
drwxrwxr-x 3 root 1001 3 Jan  6 11:08 ..
-rw-rw-r-- 1 root 1001 6 Jan  6 11:09 from_ubuntuserver_nfs.txt

Then from the FreeNAS shell the TXT file does not appear:

Code:

[root@freenas] /mnt/storage/test> ls -la
total 57
drwxrwxr-x  3 root  media  3 Jan  6 11:08 ./
drwxr-xr-x  9 root  wheel  9 Jan  6 10:59 ../
drwxrwxr-x  2 root  media  2 Jan  6 11:00 testchild/
[root@freenas] /mnt/storage/test> cd testchild/
[root@freenas] /mnt/storage/test/testchild> ls -la
total 38
drwxrwxr-x  2 root  media  2 Jan  6 11:00 ./
drwxrwxr-x  3 root  media  3 Jan  6 11:08 ../

I also copied a 'large' file into the 'testchild' folder via NFS to see if the volume used GUI updated - it did not.

Now if I delete the 'testchild' data set, and repeat the check from the FreeNAS shell, i get:

Code:

[root@freenas] /mnt/storage/test> ls -la
total 57
drwxrwxr-x  3 root  media  3 Jan  6 11:08 ./
drwxr-xr-x  9 root  wheel  9 Jan  6 10:59 ../
drwxr-xr-x  2 root  media  3 Jan  6 11:09 testchild/
[root@freenas] /mnt/storage/test> cd testchild
[root@freenas] /mnt/storage/test/testchild> ls -la
total 48
drwxr-xr-x  2 root  media  3 Jan  6 11:09 ./
drwxrwxr-x  3 root  media  3 Jan  6 11:08 ../
-rw-rw-r--  1 root  media  6 Jan  6 11:09 from_ubuntuserver_nfs.txt

So I think that NFS is treating 'testchild' as a folder, rather than a dataset and ZFS via shell/CIFS hides this folder and uses the dataset instead.

Which with my limited knowledge means that snapshots on these child datasets will snap empty folders.

Anybody able to comment/correct my findings? I am running FreeNAS-8.3.0-RELEASE-p1-x64 (r12825) under ESXi VM.

Mark

 

[William Grzybowski]

That is not a bug.

NFS shares does not span filesystem. A dataset is considered a new filesystem in the OS level, not a folder.

 

[jgreco]

I think he just discovered under-mountpoint content hiding, an ancient UNIX technique.

 

[gizimoto]

Indeed I did re-discover this technique - I've reworked my setup to ensure NFS only ever mounts the lowest sibling ZFS dataset and ditched NFS for CIFS where suitable.

Do you think this could be outlined a little better in the manual, I see there is a one liner mentioning this, but as a noob to FreeNAS and NFS it did not hit home fully.

 

[jgreco]

Where's the one-liner and ... well, here's kind of the problem for me. Having used UNIX for decades, this sort of "problem" doesn't even necessarily strike me so much as a problem as it does an obscure feature and potential newbie trap. So what I'm wondering is if you have a specific suggestion as to how to make it exceedingly clear. If you point me to the one-liner and provide a better short summary of the problem, I think I may have manual editing privileges and I'd be happy to change it, then maybe expand upon it if that seems prudent.

 

[Neek]

Though it's an old thread, I stumbled upon this issue tonight so I thought I'd add my $0.02:

While a ZFS dataset is, to the OS, basically a different file system, it doesn't need to be so from the perspective of NFS. What benefit does a user see in having sub-datasets not mount as subdirectories? What risks would this bring to the OS or the data?

It seems like this is one of those places where a modern filesystem layered atop a very old sharing protocol should yield a better behavior than what it actually does today.