Move kernel locking inside the sleep machinery. This enables calling

rwsleep(9) with PCATCH and rw_enter(9) with RW_INTR without the kernel
lock. In addition, now tsleep(9) with PCATCH should be safe to use
without the kernel lock if the sleep is purely time-based.

Tested by anton@, cheloha@, chris@
OK anton@, cheloha@

sys/dev/pci/drm/drm_linux.c

@@ -1,4 +1,4 @@
-/*	$OpenBSD: drm_linux.c,v 1.50 2019/10/23 10:17:40 jsg Exp $	*/
+/*	$OpenBSD: drm_linux.c,v 1.51 2019/11/30 11:19:17 visa Exp $	*/
 /*
  * Copyright (c) 2013 Jonathan Gray <jsg@openbsd.org>
  * Copyright (c) 2015, 2016 Mark Kettenis <kettenis@openbsd.org>
@@ -91,7 +91,7 @@ schedule_timeout(long timeout)
 	sleep_setup(&sls, sch_ident, sch_priority, "schto");
 	if (timeout != MAX_SCHEDULE_TIMEOUT)
 		sleep_setup_timeout(&sls, timeout);
-	sleep_setup_signal(&sls, sch_priority);
+	sleep_setup_signal(&sls);
 
 	wait = (sch_proc == curproc && timeout > 0);
 

sys/kern/kern_rwlock.c

@@ -1,4 +1,4 @@
-/*	$OpenBSD: kern_rwlock.c,v 1.43 2019/11/29 12:41:33 mpi Exp $	*/
+/*	$OpenBSD: kern_rwlock.c,v 1.44 2019/11/30 11:19:17 visa Exp $	*/
 
 /*
  * Copyright (c) 2002, 2003 Artur Grabowski <art@openbsd.org>
@@ -231,7 +231,7 @@ rw_enter(struct rwlock *rwl, int flags)
 	 */
 	unsigned int spin = (_kernel_lock_held()) ? 0 : RW_SPINS;
 #endif
-	int error;
+	int error, prio;
 #ifdef WITNESS
 	int lop_flags;
 
@@ -273,9 +273,12 @@ retry:
 		if (flags & RW_NOSLEEP)
 			return (EBUSY);
 
-		sleep_setup(&sls, rwl, op->wait_prio, rwl->rwl_name);
+		prio = op->wait_prio;
 		if (flags & RW_INTR)
-			sleep_setup_signal(&sls, op->wait_prio | PCATCH);
+			prio |= PCATCH;
+		sleep_setup(&sls, rwl, prio, rwl->rwl_name);
+		if (flags & RW_INTR)
+			sleep_setup_signal(&sls);
 
 		do_sleep = !rw_cas(&rwl->rwl_owner, o, set);
 

sys/kern/kern_synch.c

@@ -1,4 +1,4 @@
-/*	$OpenBSD: kern_synch.c,v 1.154 2019/11/12 04:20:21 visa Exp $	*/
+/*	$OpenBSD: kern_synch.c,v 1.155 2019/11/30 11:19:17 visa Exp $	*/
 /*	$NetBSD: kern_synch.c,v 1.37 1996/04/22 01:38:37 christos Exp $	*/
 
 /*
@@ -147,7 +147,7 @@ tsleep(const volatile void *ident, int priority, const char *wmesg, int timo)
 
 	sleep_setup(&sls, ident, priority, wmesg);
 	sleep_setup_timeout(&sls, timo);
-	sleep_setup_signal(&sls, priority);
+	sleep_setup_signal(&sls);
 
 	return sleep_finish_all(&sls, 1);
 }
@@ -207,6 +207,9 @@ msleep(const volatile void *ident, struct mutex *mtx, int priority,
 	KASSERT((priority & ~(PRIMASK | PCATCH | PNORELOCK)) == 0);
 	KASSERT(mtx != NULL);
 
+	if (priority & PCATCH)
+		KERNEL_ASSERT_LOCKED();
+
 	if (cold || panicstr) {
 		/*
 		 * After a panic, or during autoconfiguration,
@@ -233,7 +236,7 @@ msleep(const volatile void *ident, struct mutex *mtx, int priority,
 
 	sleep_setup(&sls, ident, priority, wmesg);
 	sleep_setup_timeout(&sls, timo);
-	sleep_setup_signal(&sls, priority);
+	sleep_setup_signal(&sls);
 
 	/* XXX - We need to make sure that the mutex doesn't
 	 * unblock splsched. This can be made a bit more
@@ -293,7 +296,7 @@ rwsleep(const volatile void *ident, struct rwlock *rwl, int priority,
 
 	sleep_setup(&sls, ident, priority, wmesg);
 	sleep_setup_timeout(&sls, timo);
-	sleep_setup_signal(&sls, priority);
+	sleep_setup_signal(&sls);
 
 	rw_exit(rwl);
 
@@ -342,11 +345,22 @@ sleep_setup(struct sleep_state *sls, const volatile void *ident, int prio,
 		panic("tsleep: not SONPROC");
 #endif
 
-	sls->sls_catch = 0;
+	sls->sls_catch = prio & PCATCH;
 	sls->sls_do_sleep = 1;
+	sls->sls_locked = 0;
 	sls->sls_sig = 1;
 	sls->sls_timeout = 0;
 
+	/*
+	 * The kernel has to be locked for signal processing.
+	 * This is done here and not in sleep_setup_signal() because
+	 * KERNEL_LOCK() has to be taken before SCHED_LOCK().
+	 */
+	if (sls->sls_catch != 0) {
+		KERNEL_LOCK();
+		sls->sls_locked = 1;
+	}
+
 	SCHED_LOCK(sls->sls_s);
 
 	p->p_wchan = ident;
@@ -417,13 +431,16 @@ sleep_finish_timeout(struct sleep_state *sls)
 }
 
 void
-sleep_setup_signal(struct sleep_state *sls, int prio)
+sleep_setup_signal(struct sleep_state *sls)
 {
 	struct proc *p = curproc;
 
-	if ((sls->sls_catch = (prio & PCATCH)) == 0)
+	if (sls->sls_catch == 0)
 		return;
 
+	/* sleep_setup() has locked the kernel. */
+	KERNEL_ASSERT_LOCKED();
+
 	/*
 	 * We put ourselves on the sleep queue and start our timeout
 	 * before calling CURSIG, as we could stop there, and a wakeup
@@ -449,20 +466,26 @@ int
 sleep_finish_signal(struct sleep_state *sls)
 {
 	struct proc *p = curproc;
-	int error;
+	int error = 0;
 
 	if (sls->sls_catch != 0) {
-		if ((error = single_thread_check(p, 1)))
-			return (error);
-		if (sls->sls_sig != 0 || (sls->sls_sig = CURSIG(p)) != 0) {
+		KERNEL_ASSERT_LOCKED();
+
+		error = single_thread_check(p, 1);
+		if (error == 0 &&
+		    (sls->sls_sig != 0 || (sls->sls_sig = CURSIG(p)) != 0)) {
 			if (p->p_p->ps_sigacts->ps_sigintr &
 			    sigmask(sls->sls_sig))
-				return (EINTR);
-			return (ERESTART);
+				error = EINTR;
+			else
+				error = ERESTART;
 		}
 	}
 
-	return (0);
+	if (sls->sls_locked)
+		KERNEL_UNLOCK();
+
+	return (error);
 }
 
 /*

sys/sys/proc.h

@@ -1,4 +1,4 @@
-/*	$OpenBSD: proc.h,v 1.281 2019/11/29 21:32:04 guenther Exp $	*/
+/*	$OpenBSD: proc.h,v 1.282 2019/11/30 11:19:17 visa Exp $	*/
 /*	$NetBSD: proc.h,v 1.44 1996/04/22 01:23:21 christos Exp $	*/
 
 /*-
@@ -611,6 +611,7 @@ struct sleep_state {
 	int sls_s;
 	int sls_catch;
 	int sls_do_sleep;
+	int sls_locked;
 	int sls_sig;
 	int sls_timeout;
 };

sys/sys/systm.h

@@ -1,4 +1,4 @@
-/*	$OpenBSD: systm.h,v 1.143 2019/11/02 16:56:18 cheloha Exp $	*/
+/*	$OpenBSD: systm.h,v 1.144 2019/11/30 11:19:17 visa Exp $	*/
 /*	$NetBSD: systm.h,v 1.50 1996/06/09 04:55:09 briggs Exp $	*/
 
 /*-
@@ -246,7 +246,7 @@ struct sleep_state;
 void	sleep_setup(struct sleep_state *, const volatile void *, int,
 	    const char *);
 void	sleep_setup_timeout(struct sleep_state *, int);
-void	sleep_setup_signal(struct sleep_state *, int);
+void	sleep_setup_signal(struct sleep_state *);
 void	sleep_finish(struct sleep_state *, int);
 int	sleep_finish_timeout(struct sleep_state *);
 int	sleep_finish_signal(struct sleep_state *);