Re: connection_close() called with c_struct_state == SLAP_C_INACTIVE

[Pierangelo Masarati <ando@sys-net.it>]

Howard Chu wrote:
> Pierangelo Masarati wrote:
> > Recently, with 2.3.30, I got an assertion during shutdown because 
> > connection_close() was called with c_struct_state == 
> > SLAP_C_INACTIVE.  Unfortunately I couldn't save neither the core nor 
> > a backtrace.  I guess the fix in connection.c 1.375 -> 1.376 should 
> > be backported to re23, but the code changed since then, and I fear I 
> > could break something.  Is this correct?
>
> Pretty sure that change is already incorporated in 2.3.30's connection.c.
Here's something similar, although in a different place, happening on 
Solaris 9 with a hacked version of 2.3.30 (hacking is basically not in 
this file; I only imported fixes to ITS#4769 and ITS#4771) when I shut 
it down hitting CTRL-C.

bash-2.03$ LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib/lwp/sparcv9/ 
LD_PRELOAD=/usr/lib/sparcv9/libmtmalloc.so.1:/home/sysnet/gprof-helper.so 
./bin/slapd-sysnet-0.3.34-beta.2-gprof -f slapd.conf -h ldap://:9020 -s0 -d0
pthreads: using profiling hooks for gprof
^CAssertion failed: c->c_struct_state == SLAP_C_USED, file connection.c, 
line 811
Abort (core dumped)

Note: connection.c:811 corresponds to connection.c:809 in 2.3.30 code.

(gdb) bt
#0  0x7fffffff7e7a754c in _lwp_kill () from /usr/lib/64/libc.so.1
#1  0x7fffffff7e753e7c in raise () from /usr/lib/64/libc.so.1
#2  0x7fffffff7e73d274 in abort () from /usr/lib/64/libc.so.1
#3  0x7fffffff7e73d56c in _assert () from /usr/lib/64/libc.so.1
#4  0x0000000100033834 in connection_closing (c=0x100eebb00,
   why=0x1001d0498 "slapd shutdown") at connection.c:811
#5  0x0000000100031c2c in connections_shutdown () at connection.c:212
#6  0x000000010002fd10 in slapd_daemon_task (ptr=0x100343b10) at 
daemon.c:2501
#7  0x7fffffff7f800d00 in wrapper_routine (data=0xffffffff7ffff768)
   at gprof-helper.c:81
#8  0x7fffffff7df17478 in _lwp_start ()
  from /usr/lib/lwp/sparcv9//libthread.so.1
#9  0x7fffffff7df17478 in _lwp_start ()
  from /usr/lib/lwp/sparcv9//libthread.so.1

What I see is that c->c_struct_state is 1 (SLAP_C_UNUSED) instead of 2 
(SLAP_C_USED).

Any thoughts?  I have access to the binary and the core (sort of: it's 
more than 2000 km apart on a not fast enough connection to download it, 
and the firewall will shut the connection down shortly...).

Apparently, connections_shutdown() should lock the connection's c_mutex 
before testing c_struct_state, and not after, right?

p.