misc/151111: vnodes leakage during zfs unmount

[Gleb Kurtsou]

On (18/10/2010 12:30), Gleb Smirnoff wrote:
> The following reply was made to PR kern/151111; it has been noted by GNATS.
> 
> From: Gleb Smirnoff <glebius at FreeBSD.org>
> To: "Oleg A. Mamontov" <oleg at mamontov.net>
> Cc: kib at FreeBSD.org, pjd at FreeBSD.org, freebsd-gnats-submit at FreeBSD.org
> Subject: Re: misc/151111: vnodes leakage during zfs unmount
> Date: Mon, 18 Oct 2010 16:11:55 +0400
> 
>  On Thu, Sep 30, 2010 at 11:56:18AM +0000, Oleg A. Mamontov wrote:
>  O> for i in `jot 1000`; do zfs mount tank; printf '%4d   ' $i; sysctl vfs.numvnodes; zfs umount tank; done
>  O> ########################
>  O>  1   vfs.numvnodes: 708
>  O>  2   vfs.numvnodes: 709
>  O>  3   vfs.numvnodes: 710
>  O>  4   vfs.numvnodes: 711
>  O>  5   vfs.numvnodes: 712
>  O> 
>  O> ..
>  O> 
>  O> 995   vfs.numvnodes: 1702
>  O> 996   vfs.numvnodes: 1703
>  O> 997   vfs.numvnodes: 1704
>  O> 998   vfs.numvnodes: 1705
>  O> 999   vfs.numvnodes: 1706
>  O> 1000   vfs.numvnodes: 1707

In my tests zfs vnodes are getting properly gc'ed. Most likely maxvnodes
limit is not hit during the test, thus vnodes are not reclaimed.

Try making all available vnodes used, e.g. by running
'find /usr >/dev/null' before the test. Running the test you'll see
vfs.numvnodes going up and down.

It's more visible after lowering kern.maxvnodes. I've used
kern.maxvnodes=1000. Default value on my system is ~127000.

>  Here is my lame investigation of the problem. In the zfs_domount() function
>  we've got the following code:
>  
>          /* Grab extra reference. */
>          VERIFY(VFS_ROOT(vfsp, LK_EXCLUSIVE, &vp) == 0);
>          VOP_UNLOCK(vp, 0);
>  
>  I suppose this code is expected to put an extra reference on the
>  vfsp->mnt_vnodecovered vnode. Do I mistake here? If I don't then this is
>  the source of leak.
>  
>  Debugging shows that zfs_zget(), called subsequently from zfs_root(),
>  does not find an existing znode/vnode and allocates a new one, see at the
>  end of zfs_zget() function. This vnode gots a reference and is forgotten.
>  
>  Sorry, if I am being mistaken :)
>  
>  -- 
>  Totus tuus, Glebius.
> _______________________________________________
> freebsd-fs at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-fs
> To unsubscribe, send any mail to "freebsd-fs-unsubscribe at freebsd.org"