From the mailing list:

Wonderful code Volker, you've found
an amazingly long standing bug in
the SMB1 code :-).

The teardown code for:

make test TESTS=samba3.raw.mux

does:

SMBunlink("test_dir\\*"),
and if one of the files
within test_dir is open,
we should return with a
sharing violation.

However, we have a *really*
old bug inside the wildcard
unlink code, where if do_unlink()
fails whilst deleting a file
when traversing the wildcard
match, we just do "continue"
instead of erroring out
to the caller. This is fatal
in the case of deferring a
sharing violation as it
causes a dbwatch record to
be left around with no
associated SMB1 request.

When the dbwatch timeout
fires, it causes the assert
which for developer mode is
an smb_panic, as we can't
requeue the request as
we already replied to
it...

Here's the fix (attached)
for the record. With this
patch applied on top of the
messaging changes we fail the
samba3.raw.mux test, but without
causing a panic.