diff --git a/lib/libc/sys/ptrace.2 b/lib/libc/sys/ptrace.2
--- a/lib/libc/sys/ptrace.2
+++ b/lib/libc/sys/ptrace.2
@@ -2,7 +2,7 @@
 .\"	$NetBSD: ptrace.2,v 1.2 1995/02/27 12:35:37 cgd Exp $
 .\"
 .\" This file is in the public domain.
-.Dd May 4, 2021
+.Dd May 20, 2021
 .Dt PTRACE 2
 .Os
 .Sh NAME
@@ -100,6 +100,18 @@
 All other additional signal stops use
 .Dv SIGTRAP .
 .Pp
+Normally, tracing process should detach from all traced processes
+before exiting using
+.Dv PT_DETACH
+request.
+If tracing process exits without detaching, for instance due to abnormal
+termination, the destiny of the traced processes is determined by the
+.Dv kern.kill_on_debugger_exit
+sysctl value.
+If it is set to the default value 1, such traced processes are terminated.
+If set to zero, kernel implicitly detaches tracing process, traced processes
+are continued if stopped, and then continue the execution without tracing.
+.Pp
 Each traced process has a tracing event mask.
 An event in the traced process only reports a
 signal stop if the corresponding flag is set in the tracing event mask.
diff --git a/sys/kern/kern_exit.c b/sys/kern/kern_exit.c
--- a/sys/kern/kern_exit.c
+++ b/sys/kern/kern_exit.c
@@ -504,8 +504,9 @@
 			}
 		} else {
 			/*
-			 * Traced processes are killed since their existence
-			 * means someone is screwing up.
+			 * Traced processes are killed by default
+			 * since their existence means someone is
+			 * screwing up.
 			 */
 			t = proc_realparent(q);
 			if (t == p) {
@@ -522,14 +523,20 @@
 			 * orphan link for q now while q is locked.
 			 */
 			proc_clear_orphan(q);
-			q->p_flag &= ~(P_TRACED | P_STOPPED_TRACE);
+			q->p_flag &= ~P_TRACED;
 			q->p_flag2 &= ~P2_PTRACE_FSTP;
 			q->p_ptevents = 0;
 			FOREACH_THREAD_IN_PROC(q, tdt) {
 				tdt->td_dbgflags &= ~(TDB_SUSPEND | TDB_XSIG |
 				    TDB_FSTP);
 			}
-			kern_psignal(q, SIGKILL);
+			if (kern_kill_on_dbg_exit) {
+				q->p_flag &= ~P_STOPPED_TRACE;
+				kern_psignal(q, SIGKILL);
+			} else if ((q->p_flag & (P_STOPPED_TRACE |
+			    P_STOPPED_SIG)) != 0) {
+				ptrace_unsuspend(q);
+			}
 		}
 		PROC_UNLOCK(q);
 		if (ksi != NULL)
diff --git a/sys/kern/kern_sig.c b/sys/kern/kern_sig.c
--- a/sys/kern/kern_sig.c
+++ b/sys/kern/kern_sig.c
@@ -163,6 +163,11 @@
     "Fetch sigfastblock word on each syscall entry for proper "
     "blocking semantic");
 
+int kern_kill_on_dbg_exit = 1;
+SYSCTL_INT(_kern, OID_AUTO, kill_on_debugger_exit, CTLFLAG_RWTUN,
+    &kern_kill_on_dbg_exit, 0,
+    "Kill ptraced processes when debugger exits");
+
 SYSINIT(signal, SI_SUB_P1003_1B, SI_ORDER_FIRST+3, sigqueue_start, NULL);
 
 /*
diff --git a/sys/kern/sys_process.c b/sys/kern/sys_process.c
--- a/sys/kern/sys_process.c
+++ b/sys/kern/sys_process.c
@@ -610,6 +610,19 @@
 	p->p_ptevents = PTRACE_DEFAULT;
 }
 
+void
+ptrace_unsuspend(struct proc *p)
+{
+	PROC_LOCK_ASSERT(p, MA_OWNED);
+
+	PROC_SLOCK(p);
+	p->p_flag &= ~(P_STOPPED_TRACE | P_STOPPED_SIG | P_WAITED);
+	thread_unsuspend(p);
+	PROC_SUNLOCK(p);
+	itimer_proc_continue(p);
+	kqtimer_proc_continue(p);
+}
+
 static int
 proc_can_ptrace(struct thread *td, struct proc *p)
 {
@@ -1164,12 +1177,7 @@
 		 * suspended, use PT_SUSPEND to suspend it before
 		 * continuing the process.
 		 */
-		PROC_SLOCK(p);
-		p->p_flag &= ~(P_STOPPED_TRACE | P_STOPPED_SIG | P_WAITED);
-		thread_unsuspend(p);
-		PROC_SUNLOCK(p);
-		itimer_proc_continue(p);
-		kqtimer_proc_continue(p);
+		ptrace_unsuspend(p);
 		break;
 
 	case PT_WRITE_I:
diff --git a/sys/sys/ptrace.h b/sys/sys/ptrace.h
--- a/sys/sys/ptrace.h
+++ b/sys/sys/ptrace.h
@@ -240,6 +240,9 @@
 int	proc_read_dbregs32(struct thread *_td, struct dbreg32 *_dbreg32);
 int	proc_write_dbregs32(struct thread *_td, struct dbreg32 *_dbreg32);
 #endif
+
+void	ptrace_unsuspend(struct proc *p);
+
 #else /* !_KERNEL */
 
 #include <sys/cdefs.h>
diff --git a/sys/sys/signalvar.h b/sys/sys/signalvar.h
--- a/sys/sys/signalvar.h
+++ b/sys/sys/signalvar.h
@@ -274,6 +274,7 @@
 #ifdef _KERNEL
 extern sigset_t fastblock_mask;
 extern bool sigfastblock_fetch_always;
+extern int kern_kill_on_dbg_exit;
 
 /* Return nonzero if process p has an unmasked pending signal. */
 #define	SIGPENDING(td)							\