panic: handle_written_inodeblock: bad size

[Kirk McKusick]

> Date: Wed, 21 Jul 2010 02:10:58 -0700
> From: Jeremy Chadwick <freebsd at jdc.parodius.com>
> To: Kirk McKusick <mckusick at mckusick.com>
> Cc: "Mikhail T." <mi+thun at aldan.algebra.com>, fs at freebsd.org
> Subject: Re: panic: handle_written_inodeblock: bad size
> 
> On Tue, Jul 20, 2010 at 02:47:51PM -0700, Jeremy Chadwick wrote:
> > On Tue, Jul 20, 2010 at 10:57:09AM -0700, Jeremy Chadwick wrote:
> > > On Tue, Jul 20, 2010 at 08:44:51AM -0700, Kirk McKusick wrote:
> > > > You are on the right track with getting the filesystem information.
> > > > Any place that one has an inode (say pointer ip), it is possible
> > > > to get the filesystem information using ip->i_fs->fs_fsmnt. The
> > > > mount point can also be found from the mount-point structure.
> > >
> > > [...] 
> > > The biggest problem (for me) is testing.  I have no idea how to
> > > trigger the error conditions in these functions.  I assume it varies;
> > > maybe through fsdb(8) or interactively dropping to DDB and forcing
> > > the condition.
> > > 
> > > I tend to do all of my work on this sort on a VM of FreeBSD (using
> > > VMware Workstation), but if testing on bare metal is required I
> > > have a testbed as well.
> > 
> > I've written (what I believe to be) the first stage of getting this
> > accomplished, and have been (slowly) testing each of the functions I
> > modified in src/sys/ufs/ffs/ffs_softdep.c.
> > [...] 
> > The diff so far, I think, is around 20KBytes.
> 
> I finished preliminary testing tonight.  There were two functions which
> I couldn't verify work because I couldn't get the kernel to call them no
> matter what I tried:
> 
> softdep_setup_allocext()
> request_cleanup()

The softdep_setup_allocext() function is only called when you set external 
attributes such as ACL parameters (try setfacl(1)).

The request_cleanup() function is only called when your kernel is
memory stressed from too many filesystem operations. You might get
it to trigger by using sysctl to cut debug.max_softdeps to say 5000
then try removing a tree with at least 10000 files in it.

> All the other functions I modified were tested by moving the panic()
> call near the top of the function and doing whatever was needed.
> Sometimes mounting a filesystem was all that was required to trigger it,
> other times I had to make a new filesystem + sync, or umount.  In one
> case I had to make a UFS1+SU filesystem.  It's not a completely 100%
> accurate test, hence "preliminary".  :-)

Sounds good.

> I also fixed a couple cosmetical items with the code (things not lining
> up right, some strings having ":" at the end of them when it should have
> been within the initial formatting string itself, and one function which
> used %s for "softdep" instead of just using the actual string itself).
> These were few and far between.
> 
> Below is the patch/diff for RELENG_8 I've come up with.  I only tested
> this on amd64:
> 
> http://jdc.parodius.com/freebsd/ffs_softdep.c.patch
> 
> Kirk, if you could review this I'd appreciate it.

I have grabbed your patch and will look it over in the next few days.

> -- 
> | Jeremy Chadwick                                   jdc at parodius.com |
> | Parodius Networking                       http://www.parodius.com/ |
> | UNIX Systems Administrator                  Mountain View, CA, USA |
> | Making life hard for others since 1977.              PGP: 4BD6C0CB |

Thanks for taking the time to go though all those panics!

	Kirk McKusick