On Thu, 19 Dec 2002, Dan Melomedman wrote:

> Ross Patterson wrote:
> > On Wednesday 18 December 2002 02:31 am, Jukka Marin wrote:
> > > On Wed, Dec 18, 2002 at 03:49:40PM +1100, Daniel Carosone wrote:
> > > > For those experiencing this problem - are you running softdep?
> > >
> > > No.  Is it safe?  Does it help with NFS (server, client)?  I use it in
> > > /tmp, but I haven't been brave enough to enable it elsewhere.
>
> > To sum up, has it been definitively established that softdeps are safe?  Or
> > the converse?  I'm not qualified to answer, but I'd really love to know.

On NetBSD 1.5, softdeps was found to responsible for panics under
heavy I/O. That's the "unsafe" part. By now (at least on NetBSD-1.6K,
1.6_STABLE) all such problems appear to have been fixed.

> The way I understand it is they were designed to guarantee metadata
> integrity, not data integrity. Something bad happens - your FS is not
> corrupted, but some of the data may be gone.

That's the ideal. With the write cache on the disk enabled, it doesn't
quite work that way, but then, it seems to be no worse than the
non-softdeps case. With a write back cache on the drive (or
controller), even mounting the file system "sync" isn't safe, in that
regard.

The performance boost is a side-effect, that comes from no longer
writing the metadata synchronously. NetBSD, by default, still writes
the metadata synchronously and the filedata asynchronously. I gather
the main people recommend, or use, softdeps on NetBSD is to get
roughly equivalent performance to mounting the file systems "async",
but without throwing all caution to the winds.

Frederick