inode64 readiness testing

inode64 readiness testing

[Peter Kimball]

Hi folks,

We've got some large XFS volumes that should probably be using the inode64 mount option, but aren't yet.  Before I go making irrevocable changes, I wanted to run my testing procedure by you to make sure I've actually tested what I think I tested.  These volumes will be shared via NFS, which is not your problem but seems to be a troublemaker.

I created a blank 1GB disk image, created an XFS filesystem on that image, and mounted it on a loopback device using the ino64 flag.  

I wrote a bunch of data to the filesystem (lots of small files), approximately 600MB.

At this point, I think I have a filesystem in which inodes use 64-bit addresses, even if the actual address value would fit in 32 bits.  I would expect any program that can't handle 64-bit addresses to barf when trying to access any data on the filesystem.

I then unmounted the filesystem and re-mounted it using the inode64 flag, just like it would be mounted in production.

I then verified that the programs I cared about (mostly NFS clients) could read all of the data I had written.  I also made sure they could write to the filesystem.

Since I haven't seen any read/write failures at this point, I feel I'm ready to sign off that we're ready to start using the inode64 flag.  Did I properly create files using 64-bit inodes?  Did I read from the filesystem in such a way that I would know if my readers were unable to handle 64-bit inodes?  Is there anything I should test that I haven't?

Thanks for all your hard work on this most useful project!
Peter



ps: not sure it makes a difference, this is on Centos 5.3 (2.6.18-128.el5), so I'm not entirely certain which XFS bugs/features have been folded in by the maintainers...

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: inode64 readiness testing

[Christoph Hellwig]

On Fri, Nov 18, 2011 at 12:33:16PM -0500, Peter Kimball wrote:
> I created a blank 1GB disk image, created an XFS filesystem on that image, and mounted it on a loopback device using the ino64 flag.  
> 
> I wrote a bunch of data to the filesystem (lots of small files), approximately 600MB.
> 
> At this point, I think I have a filesystem in which inodes use 64-bit addresses, even if the actual address value would fit in 32 bits.  I would expect any program that can't handle 64-bit addresses to barf when trying to access any data on the filesystem.

You will never not see 64-bit inodes on a filesystem that small ever.
Try to create a (sparse) 10TB loop image, and create some deep
directories in it.  This should create some larger inodes number for
you if you had it mounted with the inode64 flag.  You can verify that
by checking that the inode number returned from the stat systsem call
or from ls -i is larger than 32 bits.

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: inode64 readiness testing

[Peter Kimball]

On Nov 20, 2011, at 2:10 PM, Christoph Hellwig wrote:

> On Fri, Nov 18, 2011 at 12:33:16PM -0500, Peter Kimball wrote:
>> I created a blank 1GB disk image, created an XFS filesystem on that image, and mounted it on a loopback device using the ino64 flag.  
>> 
>> I wrote a bunch of data to the filesystem (lots of small files), approximately 600MB.
>> 
>> At this point, I think I have a filesystem in which inodes use 64-bit addresses, even if the actual address value would fit in 32 bits.  I would expect any program that can't handle 64-bit addresses to barf when trying to access any data on the filesystem.
> 
> You will never not see 64-bit inodes on a filesystem that small ever.
> Try to create a (sparse) 10TB loop image, and create some deep
> directories in it.  This should create some larger inodes number for
> you if you had it mounted with the inode64 flag.  You can verify that
> by checking that the inode number returned from the stat systsem call
> or from ls -i is larger than 32 bits.
> 

Thank you for that guide, Christoph.  I followed your directions and the directory tree I created included some >32-bit inode numbers so I was able to successfully test all of our NFS clients.

>From what I'd read, I thought that the ino64 mount option would do the work for me (bring 32-bit inode numbers into 64-bit range), apparently that is not the case.  This method worked great, hopefully the next person to search can find this happy thread.

Many thanks,
Peter
_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: inode64 readiness testing

[Dave Chinner]

On Mon, Nov 21, 2011 at 03:46:04PM -0500, Peter Kimball wrote:
> 
> On Nov 20, 2011, at 2:10 PM, Christoph Hellwig wrote:
> 
> > On Fri, Nov 18, 2011 at 12:33:16PM -0500, Peter Kimball wrote:
> >> I created a blank 1GB disk image, created an XFS filesystem on
> >> that image, and mounted it on a loopback device using the ino64
> >> flag.  
> >> 
> >> I wrote a bunch of data to the filesystem (lots of small
> >> files), approximately 600MB.
> >> 
> >> At this point, I think I have a filesystem in which inodes use
> >> 64-bit addresses, even if the actual address value would fit in
> >> 32 bits.  I would expect any program that can't handle 64-bit
> >> addresses to barf when trying to access any data on the
> >> filesystem.
> > 
> > You will never not see 64-bit inodes on a filesystem that small
> > ever.  Try to create a (sparse) 10TB loop image, and create some
> > deep directories in it.  This should create some larger inodes
> > number for you if you had it mounted with the inode64 flag.  You
> > can verify that by checking that the inode number returned from
> > the stat systsem call or from ls -i is larger than 32 bits.
> > 
> 
> Thank you for that guide, Christoph.  I followed your directions
> and the directory tree I created included some >32-bit inode
> numbers so I was able to successfully test all of our NFS clients.
> 
> From what I'd read, I thought that the ino64 mount option would do
> the work for me (bring 32-bit inode numbers into 64-bit range),
> apparently that is not the case.  This method worked great,
> hopefully the next person to search can find this happy thread.

The ino64 mount option does not exist any more - it got removed
quite some time ago as it was debug-only code that nobody ever
tested or verified did the right thing...

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: inode64 readiness testing

[Eric Sandeen]

On 11/18/11 11:33 AM, Peter Kimball wrote:
> Hi folks,
> 
> We've got some large XFS volumes that should probably be using the
> inode64 mount option, but aren't yet.  Before I go making irrevocable
> changes, I wanted to run my testing procedure by you to make sure
> I've actually tested what I think I tested.  These volumes will be
> shared via NFS, which is not your problem but seems to be a
> troublemaker.
> 
> I created a blank 1GB disk image, created an XFS filesystem on that
> image, and mounted it on a loopback device using the ino64 flag.
> 
> I wrote a bunch of data to the filesystem (lots of small files),
> approximately 600MB.
> 
> At this point, I think I have a filesystem in which inodes use 64-bit
> addresses, even if the actual address value would fit in 32 bits.  I
> would expect any program that can't handle 64-bit addresses to barf
> when trying to access any data on the filesystem.
> 
> I then unmounted the filesystem and re-mounted it using the inode64
> flag, just like it would be mounted in production.
> 
> I then verified that the programs I cared about (mostly NFS clients)
> could read all of the data I had written.  I also made sure they
> could write to the filesystem.
> 
> Since I haven't seen any read/write failures at this point, I feel
> I'm ready to sign off that we're ready to start using the inode64
> flag.  Did I properly create files using 64-bit inodes?  Did I read
> from the filesystem in such a way that I would know if my readers
> were unable to handle 64-bit inodes?  Is there anything I should test
> that I haven't?

You might also take a look at the script at http://sandeen.net/wordpress/?p=9,
which can look at binaries and check them for 32-bit stat() syscalls.

-Eric

> Thanks for all your hard work on this most useful project! Peter
> 
> 
> 
> ps: not sure it makes a difference, this is on Centos 5.3
> (2.6.18-128.el5), so I'm not entirely certain which XFS bugs/features
> have been folded in by the maintainers...
> 
> _______________________________________________ xfs mailing list 
> xfs@oss.sgi.com http://oss.sgi.com/mailman/listinfo/xfs
> 

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs