Running pmfs in VMs

[Zuckerman, Boris]

Matthew,

That time I was looking for a quick answer on how technically configure a VM, so a guest Linux OS would act correctly on memmap and pmfs's  physaddr would point to physical memory, rather than to a virtualized space.

You are absolutely correct about logical sharing of disk/memory spaces in clustered file systems. They know how to coordinate lock, updates and invalidate caches. 
I had some experience allowing sharing of ext3 volumes when one host had the RW access and other hosts had RO access to "stable" objects. That was based on delegating and invalidating revoked regions.

PMFS was a very attractive and easy target for "clusterizing". It did not have much of the meta cache, had well isolated locking primitives. One use case for a clustered PMFS-like solution is logical isolation of user apps and components of a distributed name space. Another is multi-host shared PM. 

It would be desirable to build a PM based locking service, so we would not have to use network for coordination.

Regards, Boris

> -----Original Message-----
> From: Matthew Wilcox [mailto:willy at linux.intel.com]
> Sent: Monday, March 24, 2014 4:58 PM
> To: Zuckerman, Boris
> Cc: linux-pmfs at lists.infradead.org
> Subject: Re: Running pmfs in VMs
> 
> On Tue, Jan 07, 2014 at 02:46:30PM +0000, Zuckerman, Boris wrote:
> > Is it possible to run pmfs in VMs (any)?
> > How do I configure VMs to allow a correct mapping of physaddr (2 VMs can see the
> same physical range)?
> 
> Hi Boris,
> 
> You never got an answer to this, but I think it's worth discussing how we can solve
> this problem in our brave new DAX-based future.  I suspect
> ext2/ext4 can both mount the same block device simultaneously in read-only mode,
> the same way that the s390 guys did originally with their XIP work.
> 
> For read-write mounts, ext4 doesn't support the shared-disc metaphor, and I don't
> think XFS does either.  Linux does have two filesystems that do support shared-discs;
> OCFS2 and GFS2.  Assuming we present the same ranges to two VMs, and the block
> driver works, both of these filesystems should do the right thing today.
> 
> We can improve them by adding DAX support, and a quick look makes it seem like it
> would be relatively straightforward to do, the same way we added support to ext4.
> 
> Would that address your use case?