Re: [PATCH] fusion: streamline ->slave_alloc/->slave_destroy

[Christoph Hellwig <hch@lst.de>]

On Mon, Oct 11, 2004 at 02:39:54PM -0400, Moore, Eric Dean wrote:
> > > > -	hd = (MPT_SCSI_HOST *)host->hostdata;
> > > > +	mptscsih_search_running_cmds(hd, target, lun);
> > > >  
> > > > -	if (hd == NULL)
> > > > +	vdev->luns[0] &= ~(1 << lun);
> > > 
> > > Ok - I'm finally getting around to testing this patch.
> > > 
> > > The driver will oops on this line because this routine is 
> > > sometimes called when vdev=NULL.  Your patch had removed the
> > > check for this NULL pointer. We probally should add it back.
> > 
> > Do you have a backtrace of such a case?  It really shouldn't happen
> > as the vdev is allocated in ->slave_alloc, and that one returns an
> > error if it couldn't allocate.
> >
> 
> 
> The problem is mptscsih_slave_destroy is being called twice. The first
> time we free hd->Targets, the second time we hit the oops.  I'm still
> investigating. Here is back trace using source level debugger.   Also
> beyond the back trace is snip of source, having the embedded printks.

the only thing I could imagine that could cause this is a race in
num_luns manipulation.  Care to check whether the oops goes away when
you turn it into an atomic_t?