Hello,
I observed a lockup while debugging a multi-lwp process with gdb. Gdb is stuck
in wait4() even though the debuggee is stopped (SSTOP). Here is my
analysis.
proc.p_nstopped counts how many lwps are stopped in the current process. When
p_nstopped reaches p_nthreads, the parent process is woken up if blocked in
wait4(2).
When the process is unstopped (proc_unstop()), LWP_WSTOP is reset for all
lwps and all LSSTOP'ed lwps transition to LSSLEEP because setrunnable() is
called.
When the first lwp of the unstopped process runs again, p_nstopped ==
p_nthreads-1 (because p_nstopped is decremented when the lwp wakes up in
tstop()).
If this lwp calls proc_stop() before the other lwps in the process get a chance
to run and there are at least 2 other lwps, p_nstopped > p_nthreads because
proc_stop() increments p_nstopped for each LSSLEEP lwp unless LWP_WSTOP is set.
As a consequence, the wake-up test at the end of proc_stop() fails and gdb
hangs.
The suggested fix uses the LWP_WSTOP bit to ensure p_nstopped is incremented
and decremented exactly once per lwp. Here is a minimal patch doing this.
There may be other related issues but I need to spend more time studying the
code to confirm or infirm my suspicions.
Cheers,
Nicolas
Index: src/sys/kern/kern_sig.c
===================================================================
--- src.orig/sys/kern/kern_sig.c 2007-08-03 20:10:24.000000000 +0200
+++ src/sys/kern/kern_sig.c 2007-08-03 20:14:45.000000000 +0200
@@ -1358,6 +1358,7 @@
*/
if (lp->lwp_flag & LWP_WSTOP) {
--p->p_nstopped;
+ lp->lwp_flag &= ~LWP_WSTOP;
} else {
if (bootverbose)
kprintf("proc_unstop: lwp %d/%d
sleeping, not stopped\n",
@@ -1370,7 +1371,6 @@
break;
}
- lp->lwp_flag &= ~LWP_WSTOP;
}
crit_exit();
}
Index: src/sys/kern/kern_synch.c
===================================================================
--- src.orig/sys/kern/kern_synch.c 2007-08-03 20:10:24.000000000 +0200
+++ src/sys/kern/kern_synch.c 2007-08-03 20:15:46.000000000 +0200
@@ -1013,6 +1013,7 @@
}
tsleep(lp->lwp_proc, 0, "stop", 0);
p->p_nstopped--;
+ lp->lwp_flag &= ~LWP_WSTOP;
crit_exit();
}
