From the submitter:

  - this causes POSIX locking to use the thread group leader
   (p->p_leader) as the locking thread for all advisory locks.
   In non-kernel-threaded code p->p_leader == p, so this will have
   no effect.

   This results in (more) correct POSIX threaded flock-ing semantics.

   It also prevents the leader from exiting before any of the children.
   (so that p->p_leader will never be stale) in exit1().

   We have been running this patch for over a month now in our lab
   under load and at customer sites.

Submitted by:	John Plevyak <jplevyak@inktomi.com>

2 files changed, 18 insertions, 21 deletions

diff --git a/sys/kern/kern_descrip.c b/sys/kern/kern_descrip.c
index cc96d7030502..7b6ebe0f7261 100644
--- a/sys/kern/kern_descrip.c
+++ b/sys/kern/kern_descrip.c
@@ -36,7 +36,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)kern_descrip.c	8.6 (Berkeley) 4/19/94
- * $Id: kern_descrip.c,v 1.62 1999/05/30 16:52:54 phk Exp $
+ * $Id: kern_descrip.c,v 1.63 1999/05/31 11:27:30 phk Exp $
  */
 
 #include "opt_compat.h"
@@ -304,16 +304,16 @@ fcntl(p, uap)
 			if ((fp->f_flag & FREAD) == 0)
 				return (EBADF);
 			p->p_flag |= P_ADVLOCK;
-			return (VOP_ADVLOCK(vp, (caddr_t)p, F_SETLK, &fl, flg));
+			return (VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_SETLK, &fl, flg));
 
 		case F_WRLCK:
 			if ((fp->f_flag & FWRITE) == 0)
 				return (EBADF);
 			p->p_flag |= P_ADVLOCK;
-			return (VOP_ADVLOCK(vp, (caddr_t)p, F_SETLK, &fl, flg));
+			return (VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_SETLK, &fl, flg));
 
 		case F_UNLCK:
-			return (VOP_ADVLOCK(vp, (caddr_t)p, F_UNLCK, &fl,
+			return (VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_UNLCK, &fl,
 				F_POSIX));
 
 		default:
@@ -334,7 +334,7 @@ fcntl(p, uap)
 			return (EINVAL);
 		if (fl.l_whence == SEEK_CUR)
 			fl.l_start += fp->f_offset;
-		if ((error = VOP_ADVLOCK(vp,(caddr_t)p,F_GETLK,&fl,F_POSIX)))
+		if ((error = VOP_ADVLOCK(vp,(caddr_t)p->p_leader,F_GETLK,&fl,F_POSIX)))
 			return (error);
 		return (copyout((caddr_t)&fl, (caddr_t)(intptr_t)uap->arg,
 		    sizeof(fl)));
@@ -1064,7 +1064,7 @@ closef(fp, p)
 		lf.l_len = 0;
 		lf.l_type = F_UNLCK;
 		vp = (struct vnode *)fp->f_data;
-		(void) VOP_ADVLOCK(vp, (caddr_t)p, F_UNLCK, &lf, F_POSIX);
+		(void) VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_UNLCK, &lf, F_POSIX);
 	}
 	if (--fp->f_count > 0)
 		return (0);
diff --git a/sys/kern/kern_exit.c b/sys/kern/kern_exit.c
index fd66be8cad01..91a746182aba 100644
--- a/sys/kern/kern_exit.c
+++ b/sys/kern/kern_exit.c
@@ -36,7 +36,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)kern_exit.c	8.7 (Berkeley) 2/12/94
- * $Id: kern_exit.c,v 1.79 1999/04/28 01:04:26 luoqi Exp $
+ * $Id: kern_exit.c,v 1.80 1999/04/28 11:36:52 phk Exp $
  */
 
 #include "opt_compat.h"
@@ -143,21 +143,10 @@ exit1(p, rv)
 			kill(p, &killArgs);
 			nq = q;
 			q = q->p_peers;
-			/*
-			 * orphan the threads so we don't mess up
-			 * when they call exit
-			 */
-			nq->p_peers = 0;
-			nq->p_leader = nq;
 		}
-
-	/* otherwise are we a peer? */
-	} else if(p->p_peers) {
-		q = p->p_leader;
-		while(q->p_peers != p)
-			q = q->p_peers;
-		q->p_peers = p->p_peers;
-	}
+		while (p->p_peers) 
+		  tsleep((caddr_t)p, PWAIT, "exit1", 0);
+	} 
 
 #ifdef PGINPROF
 	vmsizmon();
@@ -200,6 +189,14 @@ exit1(p, rv)
 	 */
 	fdfree(p);
 
+	if(p->p_leader->p_peers) {
+		q = p->p_leader;
+		while(q->p_peers != p)
+			q = q->p_peers;
+		q->p_peers = p->p_peers;
+		wakeup((caddr_t)p->p_leader);
+	}
+
 	/*
 	 * XXX Shutdown SYSV semaphores
 	 */