Re: [PATCH -V14 0/11] Generic name to handle and open by handle syscalls

["Aneesh Kumar K. V" <aneesh.kumar@linux.vnet.ibm.com>]

On Fri, 9 Jul 2010 12:42:42 -0600, Andreas Dilger <adilger@dilger.ca> wrote:
> On 2010-07-08, at 06:21, Neil Brown wrote:
> > On Thu, 08 Jul 2010 16:10:09 +0530
> > "Aneesh Kumar K. V" <aneesh.kumar@linux.vnet.ibm.com> wrote:
> >> How about adding mnt_id to the handle ? Documentation file says it is unique 
> >> 
> >> (1) mount ID:  unique identifier of the mount (may be reused after umount)
> 
> But this value is not persistent across a reboot, or even an
> umount/mount so it is not useful as an identifier.


mount id should not be looked at as a persistent identifier. It should
be used to derive a persistent identifier from /proc/self/mountinfo. The
persistent identifier could be the combination of device properties,
file system properties or the uuid which is going to be an optional
tag in /proc/self/mountinfo. 

This also implies we need to hold a reference in the mount to make sure
we can safely lookup uuid using mount id.

> 
> I suppose one way to resolve this issue is to just allow the
> underlying filesystem to supply a completely opaque filehandle to
> userspace.  For local filesystems that don't care about persistence or
> uniqueness between nodes they can use something like mount_id, and for
> distributed/clustered filesystems they can include a globally-unique
> identifier.


We could use mountid to get the persistent id from mountinfo right ? So
file handle request would include

fd = open(name);
file_handle = fd_to_handle(fd);
fs_uuid = get_uuid(file_handle.mnt_id);
close(fd);

So for your usecase the handle send to other nodes include will include
cluster_fs_uuid and file_identifier. 

-aneesh