Re: 8.0-CURRENT kernel panic on system boot

On Thu, Aug 21, 2008 at 09:33:55PM +0800, kevin wrote:
> Hi,
>    I update to the newest source tree and rebuild kernel(with zfs 
> patch).kernel panic on system startup.Should I rebuild world?
> ......
> Starting gdm.
> Sleeping thread (tid 100068,pid 1145) owns a non-sleepable lock
> sched_switch() at sched_switch+0x184
> mi_switch() at mi_switch+0x21b
> sleepq_switch() at sleepq_switch+0xfa
> sleepq_wait() at sleepq_wait+0x46
> _sx_slock_hard() at _sx_slock_hard+0x169
> _sx_slock() at _sx_slock+0x6b
> acct_process() at acct_process+0x24b
> exit1() at exit1+0x3e8
> sys_exit() at sys_exit+0xe
> syscall() at syscall+0x1dd
> Xfast_syscall() at Xfast_syscall+0xab
> --- syscall (1, FreeBSD ELF64, sys_exit), rip = 0x80094dd0c, rsp = 
> 0x7fffffffcdf8, rbp = 0x62d548 ---
> panic: sleeping thread
> cpuid = 0
> KDB: enter: panic
> [thread pid 1143 tid 100127 ]
> Stopped at       kdb_enter+0x3d: movq    $0,0x6cebb4(%rip)
> db>bt
> Tracing pid 1143 tid 100127 td 0xfffff000f073700
> kdb_enter() at kdb_enter()+0x3d
> panic() at panic()+0x16c
> propagate_priority() propagate_priority+0x221
> turnstile_wait() at turnstile_wait+0x212
> _mtx_lock_sleep() at _mtx_lock_sleep+0xe0
> _mtx_lock_flags() at _mtx_lock_flags+0xc0
> kern_wait() at kern_wait+0x11b
> wait4() at wait4+0x33
> syscall() at syscall()+0x1dd
> Xfast_syscall() at Xfast_syscall()+0xab
> --- syscall (7, FreeBSD ELF64, wait4), rip = 0x8009115c, rsp = 
> 0x7fffffffcf18, rbp = 0x26 ---
> db>show locks
> exclusive sx proctree (proctree) r = 0 (0xffffffff8099fc80) locked 
> _at_/usr/src/sys/kern/kern_exit.c:702
> db>show allpcpu
> Current CPU:0
> cpuid   = 0
> curthread =0xffffff000f073700:pid 1143 "sh"
> curpcb    =0xfffffffea6084d40
> fpcurthread  = none
> idlethread  = 0xffffff0002395380:pid 11 "idle: cpu0"
> spin locks held:
> 
> cpuid   = 1
> curthread =0xffffff0002395700:pid 1143 "idle: cpu1"
> curpcb    =0xfffffffe4001dd40
> fpcurthread  = none
> idlethread  = 0xffffff0002395700:pid 11 "idle: cpu1"
> spin locks held:
> 

No, this is a consequence of taking sx after proc lock in kern_acct.c,
acct_process(). You may turn off accounting, or try this patch:

diff --git a/sys/kern/kern_acct.c b/sys/kern/kern_acct.c
index e505f62..d45e3b7 100644
--- a/sys/kern/kern_acct.c
+++ b/sys/kern/kern_acct.c
_at__at_ -366,16 +366,25 _at__at_ acct_process(struct thread *td)
 	 * Get process accounting information.
 	 */
 
+	sx_slock(&proctree_lock);
 	PROC_LOCK(p);
-	/* (1) The name of the command that ran */
+
+	/* (1) The terminal from which the process was started */
+	if ((p->p_flag & P_CONTROLT) && p->p_pgrp->pg_session->s_ttyp)
+		acct.ac_tty = tty_udev(p->p_pgrp->pg_session->s_ttyp);
+	else
+		acct.ac_tty = NODEV;
+	sx_sunlock(&proctree_lock);
+
+	/* (2) The name of the command that ran */
 	bcopy(p->p_comm, acct.ac_comm, sizeof acct.ac_comm);
 
-	/* (2) The amount of user and system time that was used */
+	/* (3) The amount of user and system time that was used */
 	rufetchcalc(p, &ru, &ut, &st);
 	acct.ac_utime = encode_timeval(ut);
 	acct.ac_stime = encode_timeval(st);
 
-	/* (3) The elapsed time the command ran (and its starting time) */
+	/* (4) The elapsed time the command ran (and its starting time) */
 	tmp = boottime;
 	timevaladd(&tmp, &p->p_stats->p_start);
 	acct.ac_btime = tmp.tv_sec;
_at__at_ -383,7 +392,7 _at__at_ acct_process(struct thread *td)
 	timevalsub(&tmp, &p->p_stats->p_start);
 	acct.ac_etime = encode_timeval(tmp);
 
-	/* (4) The average amount of memory used */
+	/* (5) The average amount of memory used */
 	tmp = ut;
 	timevaladd(&tmp, &st);
 	/* Convert tmp (i.e. u + s) into hz units to match ru_i*. */
_at__at_ -394,21 +403,13 _at__at_ acct_process(struct thread *td)
 	else
 		acct.ac_mem = 0;
 
-	/* (5) The number of disk I/O operations done */
+	/* (6) The number of disk I/O operations done */
 	acct.ac_io = encode_long(ru.ru_inblock + ru.ru_oublock);
 
-	/* (6) The UID and GID of the process */
+	/* (7) The UID and GID of the process */
 	acct.ac_uid = p->p_ucred->cr_ruid;
 	acct.ac_gid = p->p_ucred->cr_rgid;
 
-	/* (7) The terminal from which the process was started */
-	sx_slock(&proctree_lock);
-	if ((p->p_flag & P_CONTROLT) && p->p_pgrp->pg_session->s_ttyp)
-		acct.ac_tty = tty_udev(p->p_pgrp->pg_session->s_ttyp);
-	else
-		acct.ac_tty = NODEV;
-	sx_sunlock(&proctree_lock);
-
 	/* (8) The boolean flags that tell how the process terminated, etc. */
 	acct.ac_flagx = p->p_acflag;
 	PROC_UNLOCK(p);