Re: [Qemu-devel] Re: file system sharing

[Mark Williamson]

> Mark Williamson wrote:
> > If only one machine (host or guest) has mounted the device then it should
> > always be safe to do this.  You may get away with read only mounting in
> > one and writing in the other but it's not a reliable solution.  Never
> > allow more than one writer to the filesystem - this does bad things to
> > your filesystem!
>
> I am well aware of these issues having worked on filesystems in the
> 1980s and having designed and built one of the first RPC based networked
> filesystems in the late 80s for a CAD/CAM company.  To this day the nova
> architecture gives me the willies.

Ah yes.  Somehow whenever I work on a substantial project, I'm left with a new 
phobia ;-)  Just thought I'd mention it - I've hosed my own filesystems in 
similar ways (admittedly I should have thought what I was doing).

> I suspect based on a comment that someone else made about caching that I
> would need some sort of event to trigger a flush either by an explicit
> flush call or a close.  Would unmounting the disk image create such an
> event?

As far as QEmu's block layer is concerned?  Good question - I don't know 
enough about its internals to comment.  Can anyone verify if QEmu does 
internal caching for block IOs? (or anything else that'd cause coherency 
problems here, e.g. async IOs seem a likely candidate).

> > If you're using a file-based disk and it's partitioned you'll need to use
> > lomount
> > http://www.dad-answers.com/qemu/utilities/QEMU-HD-Mounter/lomount/ to
> > mount the right partition in the host.
>
> so what I can do is create the partition image on the host, start up
> qemu with that multi-partition disk image, do what I need to in qemu,
> shut down qemu and then I have a modified disk image.  if I need to
> modify it from the host, then I can use lomount to make it accessible as
> the local filesystem.

Yep.  You *could* take a disk file and format it all with one filesystem, 
which would make it directly mountable under the guest Linux.  However, the 
Linux loop driver doesn't like /partitioned/ "file disks" - lomount sorts 
this out in userspace by just mounting one partition at a time.  My thought 
was that if you were using (say) an image of a USB key, it'd probably have a 
partition table on it.

> > I imagine just giving the guest access to the device file would work.
>
> I'm not entirely sure how to do that.  would I just do something like:
>
> -hdc /dev/sda1
>
> as part of the command line?

That might work.  Alternatively, exporting the whole device -hdc /dev/sda 
might be worth a try.  Can't say I've ever tried it (so far!)

The alternative would be to use some network-based solution.  I'm sure you've 
thought of that but I seem to remember a user-friendly way of letting a QEmu 
guest access an SMB share, which would make it rather more convenient to set 
up.

Cheers,
Mark