svn commit: r294329 - in head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs: . sys

[Slawa Olhovchenkov]

On Thu, Jan 21, 2016 at 01:34:57AM +0200, Andriy Gapon wrote:

> On 20/01/2016 22:03, Alan Somers wrote:
> > On Wed, Jan 20, 2016 at 2:20 AM, Andriy Gapon <avg at freebsd.org> wrote:
> >> On 19/01/2016 19:20, Alan Somers wrote:
> >>> The thing is, it never really worked in the first place.  Panics and
> >>> deadlocks are so frequent that I don't think the feature was usable
> >>> for anybody.
> >>
> >> The feature is perfectly usable for me.  I have never run into the problems that
> >> you describe.  Why not fix the real bugs that you've run into?
> > 
> > Spectra Logic and iXSystems both experienced many problems with this.
> > The worst is a deadlock that can be triggered simply by pulling a
> > drive from a redundant pool when there exists a zvol anywhere in the
> > system (see https://reviews.freebsd.org/D4998 for a quick way to
> > reproduce).  Fixing it correctly would likely require far more time
> > than I have available.  I just want the bugs to go away.  See that
> > same code review for a change to make the feature optional.
> 
> I think that we all want all the bugs to go way.  One way to remove bugs is to
> remove (disable) code that contains bugs.  That way the perfect bug-free
> software is clearly achievable :-)  Unfortunately, that technique is not always
> welcomed.
> 
> P.S.
> I think that the real problem here is that a method of a geom must never drop
> topology_lock.  In other words, the GEOM management code (like g_xxx() stuff in
> geom_subr.c) expects that a topology can not change underneath it.  But
> zvol_geom_access() clearly breaks that contract.

May be same cause problem with swap on zvol (don't test on latest
-stable)?