UFS2 with Soft Updates Robust?

[Erik Trulsson]

On Thu, Dec 23, 2004 at 08:53:18AM -0900, Damien Hull wrote:
> On Thu, 2004-12-23 at 09:52 +0100, Erik Trulsson wrote:
> > On Thu, Dec 23, 2004 at 06:48:41AM -0000, John Conover wrote:
> > > Erik Trulsson writes:
> > > > On Wed, Dec 22, 2004 at 08:57:00PM -0900, Damien Hull wrote:
> > > > > On Tue, 2004-12-21 at 21:31 -0500, Chuck Swiger wrote:
> > > > > > John Conover wrote:
> > > > > > > Is UFS2 with soft updates the most robust file system in freebsd?
> > > > > > 
> > > > > > No, although UFS2 with softupdates is robust enough for production use.
> > > > > > 
> > > > > > If you make the filesystem writes syncronous and disable write caching on the 
> > > > > > hard drive, you will improve the robustness at significant cost to performance.
> > > > > > 
> > > > > 
> > > > > Are you saying that the UFS2 file system sucks?
> > > > 
> > > > Not at all, but standard IDE-drives suck when it comes to robustness.
> > > > (They tend to lie and tell the OS that data has been written to the
> > > > disk, when in reality it has only been written to the disks cache.)
> > > > (Thus the advice above to turn off write-caching for maximum
> > > > robustness.)
> > > > 
> > > > If you use softupdates (on a disk that doesn't lie) the filesystem on
> > > > the disk will always be consistent, but data written during the last 30
> > > > seconds or so might not yet have been written to the disk, and can therefore
> > > > be lost if e.g. the power to the computer is turned off.
> > > >
> > > 
> > > Erik, does that mean if you use softupdates, (on a SCSI,) that,
> > > although file(s) currently being written may be truncated since the
> > > cache is not flushed, that the file system can be repaired
> > > automatically by fsck to a consistent state?
> > 
> > That is the idea.  With softupdates the filesystem *on the disk* should
> > always be in a consistent state, such that fsck should always succeed.
> > (At least that is the theory, I haven't tested how well it holds in
> > reality.)
> > 
> > 
> > > 
> > > Even without synchronous writes or enabling cache write through?
> > > 
> > > 	John
> > > 
> > > BTW, the reason for the question is that most SCSIs today have many
> > > meg of HW cache, and many, (maybe most,) controllers don't permit
> > > write through anymore. So, even if the OS flushes its cache, the HW
> > > cache may not be written to the disk-so synchronous writes and OS
> > > cache write through may be of little value.
> > 
> > The only cache that can mess up things is that on the disk itself.
> > Forget about any caches managed by the OS - they are handled by the
> > filesystem code and flushed when needed.
> > 
> > 
> > The cache on the disk is another matter.  One difference between SCSI
> > and IDE is that SCSI support tagging which lets the OS determine if a
> > given block has actually been written to the disk or just to the disk's
> > cache.  This is needed for softupdates to work correctly.
> > 
> > On IDE-disks this information is often not available which means that
> > you may need to disable the write-cache on the disk (or at least make
> > it write-through) to be able to rely on the guarantees of softupdates.
> > 
> > 
> > A few IDE-disks and some of the newer S-ATA disks also support tagged
> > queueing, but support for that is not nearly as good as it is for SCSI
> > - not yet anyway.
> > 
> > 
> 
> Are you saying that it's not safe to use softupdates on IDE drives?

I am saying that it is not safe to use IDE-drive with write-caching
enabled.  This holds regardless of whether you use soft-updates or not
and for all operating systems and filesystems.

> 
> I've always used softupdates on IDE and never had a problem. However,
> I've never had a system that was under heave load. If lots of people are
> writing to the drive ( I'm thinking file server) under softupdates it
> sounds like something could go wrong. 

Something could go wrong, yes, but the risk is fairly small and the
performance penalty when you disable write-caching is large enough that
most people consider it a risk worth taking ( I certainly do.)

> 
> Lets say I have a small network of 25 users. Were going to install a
> file and print sever using FreeBSD 5.3 (UFS2). Because we don't have a
> lot of money we get SATA drives. Should I use softupdates or not? 

Using softupdates should not hurt anything compared to not using
softupdates.

> 
> If I had more then 25 users I would consider getting SCSI.
> -- 
> Damien Hull <dhull at digitaloverload.net>
> 

-- 
<Insert your favourite quote here.>
Erik Trulsson
ertr1013 at student.uu.se