freebsd stable - Aug 2003 - scsi-da does not work with INVARIANTS (fwd)

Ken is aware of the following problem.  It is in both cd(4) and da(4) as
well as stable and current.  One possible approach would be to run
{da,cd}register() from a task queue and not at interrupt time.

-Nate

---------- Forwarded message ----------
Date: Mon, 25 Aug 2003 15:45:22 -0500
From: Aniruddha Bohra <bohra@cs.rutgers.edu>
To: njl@freebsd.org
Subject: scsi-da does not work with INVARIANTS

[ Also posted to -stable ]
Hello
     I tried compiling the 4-STABLE branch with INVARIANTS and
INVARIANT_SUPPORT.

     I got a panic on boot - The problem is that daregister
calls malloc with M_WAITOK in interrupt context and INVARIANTS
panics the kernel.

  $FreeBSD: src/sys/cam/scsi/scsi_da.c,v 1.42.2.40 2003/08/24 03:26:38
ken Exp $

     The problem is in the following lines :

     softc->sysctl_tree = SYSCTL_ADD_NODE(&softc->sysctl_ctx,
     SYSCTL_STATIC_CHILDREN(_kern_cam_da), OID_AUTO, tmpstr2,
         CTLFLAG_RD, 0, tmpstr);


SYSCTL_ADD_NODE is a macro defined in sys/sysctl.h as :
#define SYSCTL_ADD_NODE(ctx, parent,nbr,name,access,handler, descr)\
     sysctl_add_oid(ctx, parent, nbr, name, CTLTYPE_NODE|access,\
     0, 0, handler, "N", descr);

And sysctl_add_oid is defined in kern/kern_sysctl.c and has the following :

oidp = malloc(sizeof(struct sysctl_oid), M_SYSCTLOID, M_WAITOK);

which has a WAITOK malloc.


     Unfortunately, I cannot get a trace till tomorrow. Right now
the machine runs fine without INVARIANTS and INVARIANT_SUPPORT.


     Thanks

Aniruddha

On Mon, Aug 25, 2003 at 13:49:30 -0700, Nate Lawson
wrote:
> Ken is aware of the following problem.  It is in both cd(4) and da(4) as
> well as stable and current.  One possible approach would be to run
> {da,cd}register() from a task queue and not at interrupt time.
That would be tricky, since the peripheral registration process currently
expects a success/failure return from the peripheral constructor.

Just putting the sysctl creation in a task queue, though, would not be too
difficult.

One question I have, though, is whether task queues run in a thread context
or not.  If not, then we'll have the same problem.

Ken
-- 
Kenneth Merry
ken@kdm.org