svn commit: r250411 - in head/sys: conf kern sys

[John Baldwin]

On Friday, May 10, 2013 2:51:20 pm Marcel Moolenaar wrote:
> 
> On May 10, 2013, at 9:11 AM, John Baldwin <jhb at freebsd.org> wrote:
> 
> > On Friday, May 10, 2013 11:46:54 am Marcel Moolenaar wrote:
> >>> 
> >>> 2) vnode locks from a local filesystem that report a LOR with a "devfs"
> >>>  vnode.  Typical reports are either "ufs" -> "devfs" or in some cases
> >>>  "ufs" -> "devfs" -> "ufs".  As with 1), I would much rather tag the
> >>>  offending location than to disable all WITNESS checking on vnode locks.
> >> 
> >> With more file system types in use, this will get mixed up with the
> >> other file systems and noise you get is rather severe. It is a big
> >> problem for us at Juniper.
> > 
> > Note, it is very specific that the second lock is always "devfs".  I think
> > that points to this being isolated to a few specific places, not a generic
> > ordering problem.
> 
> Alas, that's not the case. These LORs are reported between ufs and unionfs,
> or ufs and isofs, etc. It's not just between something and devfs.

Ugh, I have only seen them with devfs so had presumed them to be more 
localized (and thus more easily targeted).  In that case your change
may be as fine-grained as we can get.  I would also like to still keep
WITNESS checking between vnode locks and other lock types, and LK_NOWITNESS
would remove that, so between your change and Attilio's approach I do
prefer yours.

> I'm not sure the only options we have are to ignore the problem
> or implement a general fix. If we set out to silence witness for
> the known false positives then it's ok to handle them on a case
> by case basis. We'll see patterns soon enough and then re-code
> the solutions in terms of those patterns. If we're lucky we see
> a single general solution, but if not, then it's fine to have a
> handful of special case. The worse we can do is not address it
> at all.

I was assuming that the reversals were far more specific, and knowing about 
other false positives like the dirhash one, I want a generic way to do more 
fine-grained marking of false positives.  If there were only a few places we 
would need to mark to fix the reversals you see, then I would prefer the 
suspend/resume approach for your case.  However, the reversals you are masking 
sound too widespread to support that.

-- 
John Baldwin