Re: current dom0 panic on domu launch

[Christoph Egger <Christoph_Egger%gmx.de@localhost>]

On Wednesday 14 October 2009 12:29:44 Manuel Bouyer wrote:
> On Wed, Oct 14, 2009 at 12:24:48PM +0200, Christoph Egger wrote:
> > On Wednesday 14 October 2009 11:55:55 Manuel Bouyer wrote:
> > > On Wed, Oct 14, 2009 at 08:05:43AM +0200, Christoph Egger wrote:
> > > > Index: sys/dev/scsipi/sd.c
> > > > ===================================================================
> > > > RCS file: /cvsroot/src/sys/dev/scsipi/sd.c,v
> > > > retrieving revision 1.289
> > > > diff -u -p -r1.289 sd.c
> > > > --- sys/dev/scsipi/sd.c 3 Aug 2009 09:40:45 -0000       1.289
> > > > +++ sys/dev/scsipi/sd.c 14 Oct 2009 05:50:21 -0000
> > > > @@ -1188,7 +1188,9 @@ sdioctl(dev_t dev, u_long cmd, void *add
> > > >                  */
> > > >                 if ((flag & FWRITE) == 0)
> > > >                         return (EBADF);
> > > > -               if (((sd->flags & SDF_DIRTY) != 0 || *(int *)addr != 
> > > > 0)) {
> > > > +               if (((sd->flags & SDF_DIRTY) != 0 ||
> > > > +                  (addr && *(int *)addr != 0)))
> > > > +               {
> > > >                         error = sd_flush(sd, 0);
> > > >                         if (error)
> > > >                                 sd->flags &= ~SDF_FLUSHING;
> > >
> > > that's probably the wrong thing to do, some cache flush operations
> > > would be ignored. Better fix the caller to provice a non-NULL, non-0
> > > int pointer.
> >
> > This is the backtrace provided by Sarton in an earlier mail:
> >
> > sdioctl() at netbsd:sdioctl+0x506
> > VOP_IOCTL() at netbsd:VOP_IOCTL+0x31
> > xbdback_do_io() at netbsd:xbdback_do_io+0xd2
> > workqueue_worker() at netbsd:workqueue_worker+0x81
> >
> > IN xbdback_do_io(), there's this line:
> >
> >                 error = VOP_IOCTL(xbdi->xbdi_vp, DIOCCACHESYNC, NULL,
> > FWRITE, kauth_cred_get());
> >
> > where NULL is explicitely passed to addr.
> > If sd.c is correct as is, then xbdback_do_io() is wrong.
>
> I guess sd.c is correct, so I'd say xbdback_do_io() needs to be
> changed to pass a non-0 int.

I think, this depends on whether VOP_IOCTL allows to pass NULL to the data
argument for the DIOCCACHESYNC command or not.
The manpage says nothing about this.

If it does, then my patch is correct. If not, then xbdback_do_io() has
to be fixed.

Christoph