Revert the NET_LOCK() and bring back pf's contention lock for release.

For the moment the NET_LOCK() is always taken by threads running under
KERNEL_LOCK().  That means it doesn't buy us anything except a possible
deadlock that we did not spot.  So make sure this doesn't happen, we'll
have plenty of time in the next release cycle to stress test it.

ok visa@

10 files changed, 62 insertions, 77 deletions

diff --git a/sys/kern/uipc_socket.c b/sys/kern/uipc_socket.c
index 1bb5874d1fa..0d427de67c9 100644
--- a/sys/kern/uipc_socket.c
+++ b/sys/kern/uipc_socket.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: uipc_socket.c,v 1.180 2017/03/13 20:18:21 claudio Exp $	*/
+/*	$OpenBSD: uipc_socket.c,v 1.181 2017/03/17 17:19:16 mpi Exp $	*/
 /*	$NetBSD: uipc_socket.c,v 1.21 1996/02/04 02:17:52 christos Exp $	*/
 
 /*
@@ -1038,12 +1038,10 @@ sorflush(struct socket *so)
 {
 	struct sockbuf *sb = &so->so_rcv;
 	struct protosw *pr = so->so_proto;
-	sa_family_t af = pr->pr_domain->dom_family;
 	struct sockbuf asb;
 
 	sb->sb_flags |= SB_NOINTR;
-	sblock(sb, M_WAITOK,
-	    (af != PF_LOCAL && af != PF_ROUTE) ? &netlock : NULL);
+	sblock(sb, M_WAITOK, NULL);
 	socantrcvmore(so);
 	sbunlock(sb);
 	asb = *sb;
diff --git a/sys/kern/uipc_socket2.c b/sys/kern/uipc_socket2.c
index 2da9a6a61d6..ab72035f63c 100644
--- a/sys/kern/uipc_socket2.c
+++ b/sys/kern/uipc_socket2.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: uipc_socket2.c,v 1.74 2017/03/13 20:18:21 claudio Exp $	*/
+/*	$OpenBSD: uipc_socket2.c,v 1.75 2017/03/17 17:19:16 mpi Exp $	*/
 /*	$NetBSD: uipc_socket2.c,v 1.11 1996/02/04 02:17:55 christos Exp $	*/
 
 /*
@@ -299,11 +299,7 @@ soassertlocked(struct socket *so)
 int
 sosleep(struct socket *so, void *ident, int prio, const char *wmesg, int timo)
 {
-	if ((so->so_proto->pr_domain->dom_family != PF_LOCAL) &&
-	    (so->so_proto->pr_domain->dom_family != PF_ROUTE)) {
-		return rwsleep(ident, &netlock, prio, wmesg, timo);
-	} else
-		return tsleep(ident, prio, wmesg, timo);
+	return tsleep(ident, prio, wmesg, timo);
 }
 
 /*
diff --git a/sys/net/if.c b/sys/net/if.c
index c690c9281a9..c34b5b067f8 100644
--- a/sys/net/if.c
+++ b/sys/net/if.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: if.c,v 1.490 2017/03/08 09:19:45 mpi Exp $	*/
+/*	$OpenBSD: if.c,v 1.491 2017/03/17 17:19:16 mpi Exp $	*/
 /*	$NetBSD: if.c,v 1.35 1996/05/07 05:26:04 thorpej Exp $	*/
 
 /*
@@ -230,12 +230,6 @@ struct taskq	*softnettq;
 struct task if_input_task_locked = TASK_INITIALIZER(if_netisr, NULL);
 
 /*
- * Serialize socket operations to ensure no new sleeping points
- * are introduced in IP output paths.
- */
-struct rwlock netlock = RWLOCK_INITIALIZER("netlock");
-
-/*
  * Network interface utility routines.
  */
 void
@@ -1152,10 +1146,7 @@ if_clone_create(const char *name, int rdomain)
 	if (ifunit(name) != NULL)
 		return (EEXIST);
 
-	/* XXXSMP breaks atomicity */
-	rw_exit_write(&netlock);
 	ret = (*ifc->ifc_create)(ifc, unit);
-	rw_enter_write(&netlock);
 
 	if (ret != 0 || (ifp = ifunit(name)) == NULL)
 		return (ret);
@@ -1197,10 +1188,7 @@ if_clone_destroy(const char *name)
 		splx(s);
 	}
 
-	/* XXXSMP breaks atomicity */
-	rw_exit_write(&netlock);
 	ret = (*ifc->ifc_destroy)(ifp);
-	rw_enter_write(&netlock);
 
 	return (ret);
 }
diff --git a/sys/net/if_pflow.c b/sys/net/if_pflow.c
index 0ef0381c3d3..5dcdea31017 100644
--- a/sys/net/if_pflow.c
+++ b/sys/net/if_pflow.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: if_pflow.c,v 1.74 2017/02/16 10:15:12 mpi Exp $	*/
+/*	$OpenBSD: if_pflow.c,v 1.75 2017/03/17 17:19:16 mpi Exp $	*/
 
 /*
  * Copyright (c) 2011 Florian Obser <florian@narrans.de>
@@ -463,12 +463,9 @@ pflowioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
 			sc->sc_gcounter=pflowstats.pflow_flows;
 			/* send templates on startup */
 			if (sc->sc_version == PFLOW_PROTO_10) {
-				/* XXXSMP breaks atomicity */
-				rw_exit_write(&netlock);
 				s = splnet();
 				pflow_sendout_ipfix_tmpl(sc);
 				splx(s);
-				rw_enter_write(&netlock);
 			}
 		} else
 			ifp->if_flags &= ~IFF_RUNNING;
@@ -508,8 +505,6 @@ pflowioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
 		    sizeof(pflowr))))
 			return (error);
 
-		/* XXXSMP breaks atomicity */
-		rw_exit_write(&netlock);
 		s = splnet();
 		error = pflow_set(sc, &pflowr);
 		splx(s);
@@ -527,7 +522,6 @@ pflowioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
 		} else
 			ifp->if_flags &= ~IFF_RUNNING;
 
-		rw_enter_write(&netlock);
 		break;
 
 	default:
diff --git a/sys/net/pf.c b/sys/net/pf.c
index 3ddcd7726d2..39f005b066b 100644
--- a/sys/net/pf.c
+++ b/sys/net/pf.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pf.c,v 1.1018 2017/03/09 05:47:28 claudio Exp $ */
+/*	$OpenBSD: pf.c,v 1.1019 2017/03/17 17:19:16 mpi Exp $ */
 
 /*
  * Copyright (c) 2001 Daniel Hartmeier
@@ -1154,20 +1154,26 @@ pf_state_export(struct pfsync_state *sp, struct pf_state *st)
 /* END state table stuff */
 
 void
-pf_purge_expired_rules(void)
+pf_purge_expired_rules(int locked)
 {
 	struct pf_rule	*r;
 
-	NET_ASSERT_LOCKED();
-
 	if (SLIST_EMPTY(&pf_rule_gcl))
 		return;
 
+	if (!locked)
+		rw_enter_write(&pf_consistency_lock);
+	else
+		rw_assert_wrlock(&pf_consistency_lock);
+
 	while ((r = SLIST_FIRST(&pf_rule_gcl)) != NULL) {
 		SLIST_REMOVE(&pf_rule_gcl, r, pf_rule, gcle);
 		KASSERT(r->rule_flag & PFRULE_EXPIRED);
 		pf_purge_rule(r);
 	}
+
+	if (!locked)
+		rw_exit_write(&pf_consistency_lock);
 }
 
 void
@@ -1188,7 +1194,7 @@ pf_purge_thread(void *v)
 		if (++nloops >= pf_default_rule.timeout[PFTM_INTERVAL]) {
 			pf_purge_expired_fragments();
 			pf_purge_expired_src_nodes(0);
-			pf_purge_expired_rules();
+			pf_purge_expired_rules(0);
 			nloops = 0;
 		}
 
@@ -1235,20 +1241,26 @@ pf_state_expires(const struct pf_state *state)
 }
 
 void
-pf_purge_expired_src_nodes(void)
+pf_purge_expired_src_nodes(int waslocked)
 {
 	struct pf_src_node		*cur, *next;
-
-	NET_ASSERT_LOCKED();
+	int				 locked = waslocked;
 
 	for (cur = RB_MIN(pf_src_tree, &tree_src_tracking); cur; cur = next) {
 	next = RB_NEXT(pf_src_tree, &tree_src_tracking, cur);
 
 		if (cur->states == 0 && cur->expire <= time_uptime) {
-			next = RB_NEXT(pf_src_tree, &tree_src_tracking, cur);
-			pf_remove_src_node(cur);
+			if (! locked) {
+				rw_enter_write(&pf_consistency_lock);
+				next = RB_NEXT(pf_src_tree,
+				    &tree_src_tracking, cur);
+				locked = 1;
+			}
 		}
 	}
+
+	if (locked && !waslocked)
+		rw_exit_write(&pf_consistency_lock);
 }
 
 void
@@ -1293,10 +1305,7 @@ pf_remove_state(struct pf_state *cur)
 	RB_REMOVE(pf_state_tree_id, &tree_id, cur);
 #if NPFLOW > 0
 	if (cur->state_flags & PFSTATE_PFLOW) {
-		/* XXXSMP breaks atomicity */
-		rw_exit_write(&netlock);
 		export_pflow(cur);
-		rw_enter_write(&netlock);
 	}
 #endif	/* NPFLOW > 0 */
 #if NPFSYNC > 0
@@ -1321,12 +1330,13 @@ pf_remove_divert_state(struct pf_state_key *sk)
 	}
 }
 
+/* callers should hold the write_lock on pf_consistency_lock */
 void
 pf_free_state(struct pf_state *cur)
 {
 	struct pf_rule_item *ri;
 
-	NET_ASSERT_LOCKED();
+	splsoftassert(IPL_SOFTNET);
 
 #if NPFSYNC > 0
 	if (pfsync_state_in_use(cur))
@@ -1361,8 +1371,7 @@ pf_purge_expired_states(u_int32_t maxcheck)
 {
 	static struct pf_state	*cur = NULL;
 	struct pf_state		*next;
-
-	NET_ASSERT_LOCKED();
+	int			 locked = 0;
 
 	while (maxcheck--) {
 		/* wrap to start of list when we hit the end */
@@ -1377,14 +1386,25 @@ pf_purge_expired_states(u_int32_t maxcheck)
 
 		if (cur->timeout == PFTM_UNLINKED) {
 			/* free removed state */
+			if (! locked) {
+				rw_enter_write(&pf_consistency_lock);
+				locked = 1;
+			}
 			pf_free_state(cur);
 		} else if (pf_state_expires(cur) <= time_uptime) {
 			/* remove and free expired state */
 			pf_remove_state(cur);
+			if (! locked) {
+				rw_enter_write(&pf_consistency_lock);
+				locked = 1;
+			}
 			pf_free_state(cur);
 		}
 		cur = next;
 	}
+
+	if (locked)
+		rw_exit_write(&pf_consistency_lock);
 }
 
 int
diff --git a/sys/net/pf_ioctl.c b/sys/net/pf_ioctl.c
index 56a43a55ab8..887a36481d6 100644
--- a/sys/net/pf_ioctl.c
+++ b/sys/net/pf_ioctl.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pf_ioctl.c,v 1.307 2017/01/30 17:41:34 benno Exp $ */
+/*	$OpenBSD: pf_ioctl.c,v 1.308 2017/03/17 17:19:16 mpi Exp $ */
 
 /*
  * Copyright (c) 2001 Daniel Hartmeier
@@ -111,6 +111,7 @@ void			 pf_qid2qname(u_int16_t, char *);
 void			 pf_qid_unref(u_int16_t);
 
 struct pf_rule		 pf_default_rule, pf_default_rule_new;
+struct rwlock		 pf_consistency_lock = RWLOCK_INITIALIZER("pfcnslk");
 
 struct {
 	char		statusif[IFNAMSIZ];
@@ -986,7 +987,12 @@ pfioctl(dev_t dev, u_long cmd, caddr_t addr, int flags, struct proc *p)
 			return (EACCES);
 		}
 
-	NET_LOCK(s);
+	if (flags & FWRITE)
+		rw_enter_write(&pf_consistency_lock);
+	else
+		rw_enter_read(&pf_consistency_lock);
+
+	s = splsoftnet();
 	switch (cmd) {
 
 	case DIOCSTART:
@@ -2382,7 +2388,11 @@ pfioctl(dev_t dev, u_long cmd, caddr_t addr, int flags, struct proc *p)
 		break;
 	}
 fail:
-	NET_UNLOCK(s);
+	splx(s);
+	if (flags & FWRITE)
+		rw_exit_write(&pf_consistency_lock);
+	else
+		rw_exit_read(&pf_consistency_lock);
 	return (error);
 }
 
diff --git a/sys/net/pf_norm.c b/sys/net/pf_norm.c
index ae60ffa2317..ce020ef96fd 100644
--- a/sys/net/pf_norm.c
+++ b/sys/net/pf_norm.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pf_norm.c,v 1.201 2017/01/30 17:41:34 benno Exp $ */
+/*	$OpenBSD: pf_norm.c,v 1.202 2017/03/17 17:19:16 mpi Exp $ */
 
 /*
  * Copyright 2001 Niels Provos <provos@citi.umich.edu>
@@ -176,8 +176,6 @@ pf_purge_expired_fragments(void)
 	struct pf_fragment	*frag;
 	int32_t			 expire;
 
-	NET_ASSERT_LOCKED();
-
 	expire = time_uptime - pf_default_rule.timeout[PFTM_FRAG];
 	while ((frag = TAILQ_LAST(&pf_fragqueue, pf_fragqueue)) != NULL) {
 		if (frag->fr_timeout > expire)
diff --git a/sys/net/pfvar.h b/sys/net/pfvar.h
index 00e9b790a91..7bf8375b450 100644
--- a/sys/net/pfvar.h
+++ b/sys/net/pfvar.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pfvar.h,v 1.449 2017/02/22 16:12:12 mpi Exp $ */
+/*	$OpenBSD: pfvar.h,v 1.450 2017/03/17 17:19:17 mpi Exp $ */
 
 /*
  * Copyright (c) 2001 Daniel Hartmeier
@@ -1614,9 +1614,9 @@ extern void			 pf_tbladdr_remove(struct pf_addr_wrap *);
 extern void			 pf_tbladdr_copyout(struct pf_addr_wrap *);
 extern void			 pf_calc_skip_steps(struct pf_rulequeue *);
 extern void			 pf_purge_thread(void *);
-extern void			 pf_purge_expired_src_nodes();
+extern void			 pf_purge_expired_src_nodes(int);
 extern void			 pf_purge_expired_states(u_int32_t);
-extern void			 pf_purge_expired_rules();
+extern void			 pf_purge_expired_rules(int);
 extern void			 pf_remove_state(struct pf_state *);
 extern void			 pf_remove_divert_state(struct pf_state_key *);
 extern void			 pf_free_state(struct pf_state *);
@@ -1790,6 +1790,7 @@ int		 pf_addr_compare(struct pf_addr *, struct pf_addr *,
 
 extern struct pf_status	pf_status;
 extern struct pool	pf_frent_pl, pf_frag_pl;
+extern struct rwlock	pf_consistency_lock;
 
 struct pf_pool_limit {
 	void		*pp;
diff --git a/sys/sys/systm.h b/sys/sys/systm.h
index e93668641aa..f382c37d72e 100644
--- a/sys/sys/systm.h
+++ b/sys/sys/systm.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: systm.h,v 1.124 2017/02/14 09:46:21 mpi Exp $	*/
+/*	$OpenBSD: systm.h,v 1.125 2017/03/17 17:19:16 mpi Exp $	*/
 /*	$NetBSD: systm.h,v 1.50 1996/06/09 04:55:09 briggs Exp $	*/
 
 /*-
@@ -293,31 +293,23 @@ int	uiomove(void *, size_t, struct uio *);
 
 #include <sys/rwlock.h>
 
-extern struct rwlock netlock;
-
 #define	NET_LOCK(s)							\
 do {									\
-	rw_enter_write(&netlock);					\
 	s = splsoftnet();						\
 } while (0)
 
 #define	NET_UNLOCK(s)							\
 do {									\
 	splx(s);							\
-	rw_exit_write(&netlock);					\
 } while (0)
 
 #define	NET_ASSERT_LOCKED()						\
 do {									\
-	if (rw_status(&netlock) != RW_WRITE)				\
-		splassert_fail(RW_WRITE, rw_status(&netlock), __func__);\
 	splsoftassert(IPL_SOFTNET);					\
 } while (0)
 
 #define	NET_ASSERT_UNLOCKED()						\
 do {									\
-	if (rw_status(&netlock) == RW_WRITE)				\
-		splassert_fail(0, rw_status(&netlock), __func__);	\
 } while (0)
 
 __returns_twice int	setjmp(label_t *);
diff --git a/sys/uvm/uvm_vnode.c b/sys/uvm/uvm_vnode.c
index e6725d6fc32..bf5664815a5 100644
--- a/sys/uvm/uvm_vnode.c
+++ b/sys/uvm/uvm_vnode.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: uvm_vnode.c,v 1.94 2017/01/25 06:15:51 mpi Exp $	*/
+/*	$OpenBSD: uvm_vnode.c,v 1.95 2017/03/17 17:19:16 mpi Exp $	*/
 /*	$NetBSD: uvm_vnode.c,v 1.36 2000/11/24 20:34:01 chs Exp $	*/
 
 /*
@@ -1176,15 +1176,6 @@ uvn_io(struct uvm_vnode *uvn, vm_page_t *pps, int npages, int flags, int rw)
 		result = vn_lock(vn, LK_EXCLUSIVE | LK_RECURSEFAIL, curproc);
 
 	if (result == 0) {
-		int netlocked = (rw_status(&netlock) == RW_WRITE);
-
-		/*
-		 * This process may already have the NET_LOCK(), if we
-		 * faulted in copyin() or copyout() in the network stack.
-		 */
-		if (netlocked)
-			rw_exit_write(&netlock);
-
 		/* NOTE: vnode now locked! */
 		if (rw == UIO_READ)
 			result = VOP_READ(vn, &uio, 0, curproc->p_ucred);
@@ -1193,9 +1184,6 @@ uvn_io(struct uvm_vnode *uvn, vm_page_t *pps, int npages, int flags, int rw)
 			    (flags & PGO_PDFREECLUST) ? IO_NOCACHE : 0,
 			    curproc->p_ucred);
 
-		if (netlocked)
-			rw_enter_write(&netlock);
-
 		if ((uvn->u_flags & UVM_VNODE_VNISLOCKED) == 0)
 			VOP_UNLOCK(vn, curproc);
 	}