panic with CTL/FC at r251897 and beyond

[Kenneth D. Merry]

On Wed, Sep 25, 2013 at 13:43:13 -0600, Kenneth D. Merry wrote:
> On Wed, Sep 25, 2013 at 03:50:51 +0000, John wrote:
> > Hi Folks,
> > 
> >    I've slowly been chasing down a panic with 9-stable and I've
> > gotten to the point where the large mfc commits of r251897 and
> > r251874 by Scott seem to be the culprit. I posted some info
> > about this a few weeks back.
> > 
> >    The last working commit (involving cam) is r251852 by Alexander.
> > This is the last commit where I can create and export a lun via
> > FC and have the client use it.
> 
> Hmm, that's not good!
> 
> >    The problem I see is in cam/ctl/scsi_ctl.c:ctlfedone() at
> > the end of the XPT_CONT_TARGET_IO: switch label:
> > 
> > /*
> >  * Release the CTIO.  The ATIO will be sent back
> >  * down to the SIM once we send status.
> >  */
> > softc->ccbs_freed++;
> > xpt_release_ccb(done_ccb);
> >                                 
> > /* Call the backend move done callback */
> > io->scsiio.be_move_done(io);
> > 
> >    be_move_done is null so the code branches to 0
> > 
> > 
> >    Before I start trying to figure out the large number of
> > changes in the above 2 commits (a pair of mass mfc's) I was
> > hoping someone might have an idea of what is wrong and could
> > provide some pointers. I'd also be curious to know if anyone
> > is successfully using a recent 9-stable with CTL/FC and if
> > so, how they have it configured.
> > 
> >    Fbsd-10 works correctly on the same hardware also.
> > 
> >    I can provide ssh access and serial console debugging
> > if needed.
> 
> Try the attached patch.  I neglected to MFC it, and you'll certainly have
> problems without it with the block backend.
> 
> The ramdisk backend should work okay without this patch, because it doesn't
> use S/G lists by default.
> 
> It's possible that the be_move_done() pointer is getting overwritten by a
> DMA.

I have confirmed that was the problem, at least for me, on stable/9.
I can create file-backed LUNs and export them via an 8Gb Qlogic FC card and
it works fine.

I ran into another bug (ctladm dumpooa causes a panic) in stable/9 that was
due to something I didn't MFC.  I just merged that fix as well.

So if you update to the latest stable/9, it should work okay.

Unfortunately this means that CTL is going to be broken in 9.2.  I emailed
RE; we'll see whether they want to merge the fixes or write it up in
the errata.

Ken
-- 
Kenneth Merry
ken at FreeBSD.ORG