> On 11/9/06, YAMAMOTO Takashi <yamt@mwd.biglobe.ne.jp> wrote:
> > > +     /* Flush the vnode if requested. */
> > > +     if (obp->b_flags & B_VFLUSH)
> > > +             vflushbuf(vp, 0);
> > > +
> > > +     /* We need to increase the number of outputs on the vnode if
> > > +      * there was any write to it (either due to a real write or due
> > > +      * to a flush). */
> > > +     if (!doread || obp->b_flags & B_VFLUSH)
> > > +             vp->v_numoutput++;
> >
> > can you explain these B_VFLUSH handling?
> 
> For the first part I just tried to flush the vnode.  I don't remember
> how I ended exactly in vflushbuf, but it *seemed* reasonable...
> Hmm... maybe one ought to call VOP_FSYNC instead?  (And maybe this
> handles the case below too.)

why do you want to flush the underlying vnode here at all?
if you want to preserve stable storage semantics or such,
it should be done regardless of B_VFLUSH.

> Aside that, the driver was getting read requests that had the B_VFLUSH
> bit active. Later on a "neg numoutput" panic happened which I was able
> to fix by incrementing v_numoutput.  Yes, that looks extremely ugly
> (at least to me) so, please, if you have any better idea on how to fix
> it tell me and I'll take a look.

i don't know how read requests with B_VFLUSH happens or how it can be fixed
by incrementing v_numoutput.  can you explain?

YAMAMOTO Takashi