Skip to contentSkip to conversation

Sparse-related updates for 3.13

PATCH In review Paul E. McKenney 3 participants 21 messages Oct 11, 2013 – Oct 13, 2013

Subsystems (8)
lore
Share & tools
lore
#0 lore raw

v3 cover Sparse-related updates for 3.13

Hello!

This series features updates to allow sparse to do a better job of statically analyzing RCU usage:

1. Add a comment indicating that despite appearances,
rcu_assign_pointer() really only evaluates its arguments once, as a cpp macro should.

2-13. Apply ACCESS_ONCE() to avoid a number of rcu_assign_pointer()
calls that would otherwise suffer sparse false positives given patch #13 below.

14. Apply ACCESS_ONCE() to rcu_assign_pointer()'s target to prevent
comiler mischief. Also require that the source pointer be from the kernel address space. Sometimes it can be from the RCU address space, which necessitates the remaining patches in this series. Which, it must be admitted, apply to a very small fraction of the rcu_assign_pointer() invocations in the kernel. This commit courtesy of Josh Triplett.

Changes from v2:

o Switch from rcu_assign_pointer() to ACCESS_ONCE() given that
the pointers are all --rcu and already visible to readers, as suggested by Eric Dumazet and Josh Triplett.

o Place the commit adding the rcu_assign_pointer()'s ACCESS_ONCE()
at the end to allow better bisectability, as suggested by Josh Triplett.

o Add a comment to rcu_assign_pointer() noting that it only evaluates
its arguments once, as suggested by Josh Triplett.

Changes from v1:

o Fix grammar nit in commit logs.

Thanx, Paul

 b/drivers/net/bonding/bond_alb.c  |    3 ++-
 b/drivers/net/bonding/bond_main.c |    8 +++++---
 b/include/linux/rcupdate.h        |   20 +++++++++++++++++++-
 b/kernel/notifier.c               |    3 ++-
 b/net/bridge/br_mdb.c             |    2 +-
 b/net/bridge/br_multicast.c       |    4 ++--
 b/net/decnet/dn_route.c           |    8 +++++---
 b/net/ipv4/ip_sockglue.c          |    3 ++-
 b/net/ipv6/ip6_gre.c              |    3 ++-
 b/net/ipv6/ip6_tunnel.c           |    3 ++-
 b/net/ipv6/sit.c                  |    3 ++-
 b/net/mac80211/sta_info.c         |    7 ++++---
 b/net/wireless/scan.c             |   32 ++++++++++++++++++--------------
 13 files changed, 66 insertions(+), 33 deletions(-)
#3 lore raw

PATCH v3 3/14 bridge: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the uses in br_multicast_del_pg() and br_multicast_new_port_group() are legitimate:
They are assigning a pointer to an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences these false positives by laundering the pointers using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

---
 net/bridge/br_multicast.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
net/bridge/br_multicast.c +2 -2
index d1c578630678..bcc4bbc16498 100644
@@ -267,7 +267,7 @@ static void br_multicast_del_pg(struct net_bridge *br,
 		if (p != pg)
 			continue;
 
-		rcu_assign_pointer(*pp, p->next);
+		ACCESS_ONCE(*pp) = p->next; /* OK: Both --rcu and visible. */
 		hlist_del_init(&p->mglist);
 		del_timer(&p->timer);
 		call_rcu_bh(&p->rcu, br_multicast_free_pg);
@@ -646,7 +646,7 @@ struct net_bridge_port_group *br_multicast_new_port_group(
 	p->addr = *group;
 	p->port = port;
 	p->state = state;
-	rcu_assign_pointer(p->next, next);
+	ACCESS_ONCE(p->next) = next; /* OK: Both --rcu and visible. */
 	hlist_add_head(&p->mglist, &port->mglist);
 	setup_timer(&p->timer, br_multicast_port_group_expired,
 		    (unsigned long)p);
Signature · 1 line
1.8.1.5
#4 lore raw

PATCH v3 4/14 wireless: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the uses in cfg80211_combine_bsses() and cfg80211_bss_update() are legitimate:
They are assigning a pointer to an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences these false positives by laundering the pointers using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

---
 net/wireless/scan.c | 32 ++++++++++++++++++--------------
 1 file changed, 18 insertions(+), 14 deletions(-)
net/wireless/scan.c +18 -14
index eeb71480f1af..ac3a47abf195 100644
@@ -670,8 +670,8 @@ static bool cfg80211_combine_bsses(struct cfg80211_registered_device *dev,
 		list_add(&bss->hidden_list, &new->hidden_list);
 		bss->pub.hidden_beacon_bss = &new->pub;
 		new->refcount += bss->refcount;
-		rcu_assign_pointer(bss->pub.beacon_ies,
-				   new->pub.beacon_ies);
+		/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+		ACCESS_ONCE(bss->pub.beacon_ies) = new->pub.beacon_ies;
 	}
 
 	return true;
@@ -705,11 +705,12 @@ cfg80211_bss_update(struct cfg80211_registered_device *dev,
 
 			old = rcu_access_pointer(found->pub.proberesp_ies);
 
-			rcu_assign_pointer(found->pub.proberesp_ies,
-					   tmp->pub.proberesp_ies);
+			/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+			ACCESS_ONCE(found->pub.proberesp_ies) =
+					   tmp->pub.proberesp_ies;
 			/* Override possible earlier Beacon frame IEs */
-			rcu_assign_pointer(found->pub.ies,
-					   tmp->pub.proberesp_ies);
+			/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+			ACCESS_ONCE(found->pub.ies) = tmp->pub.proberesp_ies;
 			if (old)
 				kfree_rcu((struct cfg80211_bss_ies *)old,
 					  rcu_head);
@@ -739,13 +740,14 @@ cfg80211_bss_update(struct cfg80211_registered_device *dev,
 
 			old = rcu_access_pointer(found->pub.beacon_ies);
 
-			rcu_assign_pointer(found->pub.beacon_ies,
-					   tmp->pub.beacon_ies);
+			/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+			ACCESS_ONCE(found->pub.beacon_ies) = tmp->pub.beacon_ies;
 
 			/* Override IEs if they were from a beacon before */
 			if (old == rcu_access_pointer(found->pub.ies))
-				rcu_assign_pointer(found->pub.ies,
-						   tmp->pub.beacon_ies);
+				/* Both --rcu & visible, ACCESS_ONCE() is OK. */
+				ACCESS_ONCE(found->pub.ies) =
+						   tmp->pub.beacon_ies;
 
 			/* Assign beacon IEs to all sub entries */
 			list_for_each_entry(bss, &found->hidden_list,
@@ -755,8 +757,9 @@ cfg80211_bss_update(struct cfg80211_registered_device *dev,
 				ies = rcu_access_pointer(bss->pub.beacon_ies);
 				WARN_ON(ies != old);
 
-				rcu_assign_pointer(bss->pub.beacon_ies,
-						   tmp->pub.beacon_ies);
+				/* Both --rcu & visible, ACCESS_ONCE() is OK. */
+				ACCESS_ONCE(bss->pub.beacon_ies) =
+						   tmp->pub.beacon_ies;
 			}
 
 			if (old)
@@ -803,8 +806,9 @@ cfg80211_bss_update(struct cfg80211_registered_device *dev,
 				list_add(&new->hidden_list,
 					 &hidden->hidden_list);
 				hidden->refcount++;
-				rcu_assign_pointer(new->pub.beacon_ies,
-						   hidden->pub.beacon_ies);
+				/* Both --rcu & visible, ACCESS_ONCE() is OK. */
+				ACCESS_ONCE(new->pub.beacon_ies) =
+						   hidden->pub.beacon_ies;
 			}
 		} else {
 			/*
Signature · 1 line
1.8.1.5
#5 lore raw

PATCH v3 5/14 decnet: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the use in dn_insert_route() is legitimate: It is assigning a pointer to an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences this false positive by laundering the pointer using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

---
 net/decnet/dn_route.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
net/decnet/dn_route.c +5 -3
index fe32388ea24f..a6ef8b025035 100644
@@ -344,8 +344,9 @@ static int dn_insert_route(struct dn_route *rt, unsigned int hash, struct dn_rou
 		if (compare_keys(&rth->fld, &rt->fld)) {
 			/* Put it first */
 			*rthp = rth->dst.dn_next;
-			rcu_assign_pointer(rth->dst.dn_next,
-					   dn_rt_hash_table[hash].chain);
+			/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+			ACCESS_ONCE(rth->dst.dn_next) =
+					   dn_rt_hash_table[hash].chain;
 			rcu_assign_pointer(dn_rt_hash_table[hash].chain, rth);
 
 			dst_use(&rth->dst, now);
@@ -358,7 +359,8 @@ static int dn_insert_route(struct dn_route *rt, unsigned int hash, struct dn_rou
 		rthp = &rth->dst.dn_next;
 	}
 
-	rcu_assign_pointer(rt->dst.dn_next, dn_rt_hash_table[hash].chain);
+	/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+	ACCESS_ONCE(rt->dst.dn_next) = dn_rt_hash_table[hash].chain;
 	rcu_assign_pointer(dn_rt_hash_table[hash].chain, rt);
 
 	dst_use(&rt->dst, now);
Signature · 1 line
1.8.1.5
#6 lore raw

PATCH v3 6/14 ipv4/ip_socketglue: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the use in ip_ra_control() is legitimate: It is assigning a pointer to an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences this false positive by laundering the pointer using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

Cc "David S. Miller" <[email protected]>
Cc Alexey Kuznetsov <[email protected]>
Cc James Morris <[email protected]>
Cc Hideaki YOSHIFUJI <[email protected]>
Cc Patrick McHardy <[email protected]>
---
 net/ipv4/ip_sockglue.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
net/ipv4/ip_sockglue.c +2 -1
index d9c4f113d709..a0e7f176e9c8 100644
@@ -269,7 +269,8 @@ int ip_ra_control(struct sock *sk, unsigned char on,
 			}
 			/* dont let ip_call_ra_chain() use sk again */
 			ra->sk = NULL;
-			rcu_assign_pointer(*rap, ra->next);
+			/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+			ACCESS_ONCE(*rap) = ra->next;
 			spin_unlock_bh(&ip_ra_lock);
 
 			if (ra->destructor)
Signature · 1 line
1.8.1.5
#7 lore raw

PATCH v3 7/14 ipv6/ip6_tunnel: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the use in ip6_tnl_unlink() is legitimate: It is assigning a pointer to an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences this false positive by laundering the pointer using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

Cc "David S. Miller" <[email protected]>
Cc Alexey Kuznetsov <[email protected]>
Cc James Morris <[email protected]>
Cc Hideaki YOSHIFUJI <[email protected]>
Cc Patrick McHardy <[email protected]>
---
 net/ipv6/ip6_tunnel.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
net/ipv6/ip6_tunnel.c +2 -1
index 61355f7f4da5..2bea7a4e49ed 100644
@@ -245,7 +245,8 @@ ip6_tnl_unlink(struct ip6_tnl_net *ip6n, struct ip6_tnl *t)
 	     (iter = rtnl_dereference(*tp)) != NULL;
 	     tp = &iter->next) {
 		if (t == iter) {
-			rcu_assign_pointer(*tp, t->next);
+			/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+			ACCESS_ONCE(*tp) = t->next;
 			break;
 		}
 	}
Signature · 1 line
1.8.1.5
#9 lore raw

PATCH v3 9/14 ipv6/sit: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the use in ipip6_tunnel_unlink() is legitimate: It is assigning a pointer to an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences this false positive by laundering the pointer using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

Cc "David S. Miller" <[email protected]>
Cc Alexey Kuznetsov <[email protected]>
Cc James Morris <[email protected]>
Cc Hideaki YOSHIFUJI <[email protected]>
Cc Patrick McHardy <[email protected]>
---
 net/ipv6/sit.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
net/ipv6/sit.c +2 -1
index 7ee5cb96db34..9b976a4b463d 100644
@@ -157,7 +157,8 @@ static void ipip6_tunnel_unlink(struct sit_net *sitn, struct ip_tunnel *t)
 	     (iter = rtnl_dereference(*tp)) != NULL;
 	     tp = &iter->next) {
 		if (t == iter) {
-			rcu_assign_pointer(*tp, t->next);
+			/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+			ACCESS_ONCE(*tp) = t->next;
 			break;
 		}
 	}
Signature · 1 line
1.8.1.5
#10 lore raw

PATCH v3 10/14 mac80211: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the uses in sta_info_hash_del() are legitimate: They are assigning a pointer to an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences this false positive by laundering the pointer using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

---
 net/mac80211/sta_info.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
net/mac80211/sta_info.c +4 -3
index aeb967a0aeed..494f03c0831f 100644
@@ -74,8 +74,8 @@ static int sta_info_hash_del(struct ieee80211_local *local,
 	if (!s)
 		return -ENOENT;
 	if (s == sta) {
-		rcu_assign_pointer(local->sta_hash[STA_HASH(sta->sta.addr)],
-				   s->hnext);
+		/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+		ACCESS_ONCE(local->sta_hash[STA_HASH(sta->sta.addr)]) = s->hnext;
 		return 0;
 	}
 
@@ -84,7 +84,8 @@ static int sta_info_hash_del(struct ieee80211_local *local,
 		s = rcu_dereference_protected(s->hnext,
 					lockdep_is_held(&local->sta_mtx));
 	if (rcu_access_pointer(s->hnext)) {
-		rcu_assign_pointer(s->hnext, sta->hnext);
+		/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+		ACCESS_ONCE(s->hnext) = sta->hnext;
 		return 0;
 	}
 
Signature · 1 line
1.8.1.5
#11 lore raw

PATCH v3 11/14 bridge/br_mdb: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the use in __br_mdb_del() is legitimate: They are assigning a pointer to an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences these false positives by laundering the pointers using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

---
 net/bridge/br_mdb.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
net/bridge/br_mdb.c +1 -1
index 85a09bb5ca51..de7197ba8695 100644
@@ -447,7 +447,7 @@ static int __br_mdb_del(struct net_bridge *br, struct br_mdb_entry *entry)
 		if (p->port->state == BR_STATE_DISABLED)
 			goto unlock;
 
-		rcu_assign_pointer(*pp, p->next);
+		ACCESS_ONCE(*pp) = p->next; /* OK: Both --rcu and visible. */
 		hlist_del_init(&p->mglist);
 		del_timer(&p->timer);
 		call_rcu_bh(&p->rcu, br_multicast_free_pg);
Signature · 1 line
1.8.1.5
#12 lore raw

PATCH v3 12/14 bonding/bond_main: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the uses in bond_change_active_slave(), bond_enslave(), and __bond_release_one() are legitimate: They are assigning a pointer to an element from an RCU-protected list (or a NULL pointer), and all elements of this list are already visible to caller.

This commit therefore silences these false positives either by laundering the pointers using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett, or by using RCU_INIT_POINTER() for NULL pointer assignments.

---
 drivers/net/bonding/bond_main.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
drivers/net/bonding/bond_main.c +5 -3
index 72df399c4ab3..e4270ae1c0a8 100644
@@ -890,7 +890,8 @@ void bond_change_active_slave(struct bonding *bond, struct slave *new_active)
 		if (new_active)
 			bond_set_slave_active_flags(new_active);
 	} else {
-		rcu_assign_pointer(bond->curr_active_slave, new_active);
+		/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+		ACCESS_ONCE(bond->curr_active_slave) = new_active;
 	}
 
 	if (bond->params.mode == BOND_MODE_ACTIVEBACKUP) {
@@ -1601,7 +1602,8 @@ int bond_enslave(struct net_device *bond_dev, struct net_device *slave_dev)
 		 * so we can change it without calling change_active_interface()
 		 */
 		if (!bond->curr_active_slave && new_slave->link == BOND_LINK_UP)
-			rcu_assign_pointer(bond->curr_active_slave, new_slave);
+			/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+			ACCESS_ONCE(bond->curr_active_slave) = new_slave;
 
 		break;
 	} /* switch(bond_mode) */
@@ -1801,7 +1803,7 @@ static int __bond_release_one(struct net_device *bond_dev,
 	}
 
 	if (all) {
-		rcu_assign_pointer(bond->curr_active_slave, NULL);
+		RCU_INIT_POINTER(bond->curr_active_slave, NULL);
 	} else if (oldcurrent == slave) {
 		/*
 		 * Note that we hold RTNL over this sequence, so there
Signature · 1 line
1.8.1.5
#8 lore raw

PATCH v3 8/14 ipv6/ip6_gre: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the use in ip6gre_tunnel_unlink() is legitimate: It is assigning a pointer to an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences this false positive by laundering the pointer using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

Cc "David S. Miller" <[email protected]>
Cc Alexey Kuznetsov <[email protected]>
Cc James Morris <[email protected]>
Cc Hideaki YOSHIFUJI <[email protected]>
Cc Patrick McHardy <[email protected]>
---
 net/ipv6/ip6_gre.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
net/ipv6/ip6_gre.c +2 -1
index 6b26e9feafb9..7bc9e1b3283e 100644
@@ -276,7 +276,8 @@ static void ip6gre_tunnel_unlink(struct ip6gre_net *ign, struct ip6_tnl *t)
 	     (iter = rtnl_dereference(*tp)) != NULL;
 	     tp = &iter->next) {
 		if (t == iter) {
-			rcu_assign_pointer(*tp, t->next);
+			/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+			ACCESS_ONCE(*tp) = t->next;
 			break;
 		}
 	}
Signature · 1 line
1.8.1.5
#13 lore raw

PATCH v3 13/14 bonding/bond_alb.c: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the use in bond_alb_handle_active_change() is legitimate: It is assigning a pointer to an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences this false positive by laundering the pointer using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

Cc "David S. Miller" <[email protected]>
Cc Alexey Kuznetsov <[email protected]>
Cc James Morris <[email protected]>
Cc Hideaki YOSHIFUJI <[email protected]>
Cc Patrick McHardy <[email protected]>
---
 drivers/net/bonding/bond_alb.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/net/bonding/bond_alb.c +2 -1
index 91f179d5135c..67d3b2893aa3 100644
@@ -1667,7 +1667,8 @@ void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave
 	}
 
 	swap_slave = bond->curr_active_slave;
-	rcu_assign_pointer(bond->curr_active_slave, new_slave);
+	/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+	ACCESS_ONCE(bond->curr_active_slave) = new_slave;
 
 	if (!new_slave || list_empty(&bond->slave_list))
 		return;
Signature · 1 line
1.8.1.5
#2 lore raw

PATCH v3 2/14 notifiers: Apply ACCESS_ONCE() to avoid sparse false positive

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The sparse checking for rcu_assign_pointer() was recently upgraded to reject non-__kernel address spaces. This also rejects __rcu, which is almost always the right thing to do. However, the use in notifier_chain_unregister() is legitimate: It is deleting an element from an RCU-protected list, and all elements of this list are already visible to caller.

This commit therefore silences this false positive by laundering the pointer using ACCESS_ONCE() as suggested by Eric Dumazet and Josh Triplett.

---
 kernel/notifier.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
kernel/notifier.c +2 -1
index 2d5cc4ccff7f..197eb70805a4 100644
@@ -51,7 +51,8 @@ static int notifier_chain_unregister(struct notifier_block **nl,
 {
 	while ((*nl) != NULL) {
 		if ((*nl) == n) {
-			rcu_assign_pointer(*nl, n->next);
+			/* Both --rcu and visible, so ACCESS_ONCE() is OK. */
+			ACCESS_ONCE(*nl) = n->next;
 			return 0;
 		}
 		nl = &((*nl)->next);
Signature · 1 line
1.8.1.5
#14 lore raw

PATCH v3 14/14 rcu: Make rcu_assign_pointer's assignment volatile and type-safe

In reply to Paul E. McKenney

From: Josh Triplett <[email protected]>

rcu_assign_pointer needs to use ACCESS_ONCE to make the assignment to the destination pointer volatile, to protect against compilers too clever for their own good.

In addition, since rcu_assign_pointer force-casts the source pointer to add the __rcu address space (overriding any existing address space), add an explicit check that the source pointer has the __kernel address space to start with.

This new check produces warnings like this, when attempting to assign from a __user pointer:

test.c:25:9: warning: incorrect type in argument 2 (different address spaces)
test.c:25:9: expected struct foo *<noident>
test.c:25:9: got struct foo [noderef] <asn:1>*badsrc

---
 include/linux/rcupdate.h | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
include/linux/rcupdate.h +11 -1
index f4da2175cde0..24206a8b8a42 100644
@@ -506,8 +506,17 @@ static inline void rcu_preempt_sleep_check(void)
 #ifdef __CHECKER__
 #define rcu_dereference_sparse(p, space) \
 	((void)(((typeof(*p) space *)p) == p))
+/* The dummy first argument in __rcu_assign_pointer_typecheck makes the
+ * typechecked pointer the second argument, matching rcu_assign_pointer itself;
+ * this avoids confusion about argument numbers in warning messages. */
+#define __rcu_assign_pointer_check_kernel(v) \
+	do { \
+		extern void __rcu_assign_pointer_typecheck(int, typeof(*(v)) __kernel *); \
+		__rcu_assign_pointer_typecheck(0, v); \
+	} while (0)
 #else /* #ifdef __CHECKER__ */
 #define rcu_dereference_sparse(p, space)
+#define __rcu_assign_pointer_check_kernel(v) do { } while (0)
 #endif /* #else #ifdef __CHECKER__ */
 
 #define __rcu_access_pointer(p, space) \
@@ -551,7 +560,8 @@ static inline void rcu_preempt_sleep_check(void)
 #define __rcu_assign_pointer(p, v, space) \
 	do { \
 		smp_wmb(); \
-		(p) = (typeof(*v) __force space *)(v); \
+		__rcu_assign_pointer_check_kernel(v); \
+		ACCESS_ONCE(p) = (typeof(*(v)) __force space *)(v); \
 	} while (0)
 
 
Signature · 1 line
1.8.1.5
#1 lore raw

PATCH v3 1/14 rcu: Add comment on evaluate-once properties of rcu_assign_pointer().

In reply to Paul E. McKenney

From: "Paul E. McKenney" <[email protected]>

The rcu_assign_pointer() macro, as with most cpp macros, must not evaluate its argument more than once. And it in fact does not. But this might not be obvious to the casual observer, because one of the arguments appears no less than three times. However, but one expansion is only visible to sparse (__CHECKER__), and one lives inside a typeof (where it will never be evaluated), so this is in fact safe.

This commit therefore adds a comment making this explicit.

---
 include/linux/rcupdate.h | 8 ++++++++
 1 file changed, 8 insertions(+)
include/linux/rcupdate.h +8 -0
index f1f1bc39346b..f4da2175cde0 100644
@@ -911,6 +911,14 @@ static inline notrace void rcu_read_unlock_sched_notrace(void)
  * rcu_assign_pointer() is a very bad thing that results in
  * impossible-to-diagnose memory corruption.  So please be careful.
  * See the RCU_INIT_POINTER() comment header for details.
+ *
+ * Note that rcu_assign_pointer() evaluates each of its arguments only
+ * once, appearances notwithstanding.  One of the "extra" evaluations
+ * is in typeof() and the other visible only to sparse (__CHECKER__),
+ * neither of which actually execute the argument.  As with most cpp
+ * macros, this execute-arguments-only-once property is important, so
+ * please be careful when making changes to rcu_assign_pointer() and the
+ * other macros that it invokes.
  */
 #define rcu_assign_pointer(p, v) \
 	__rcu_assign_pointer((p), (v), __rcu)
Signature · 1 line
1.8.1.5
maintainer
#15 lore raw
In reply to Paul E. McKenney
Quoted 5 lines · Paul E. McKenney

Changes from v2:

o Switch from rcu_assign_pointer() to ACCESS_ONCE() given that
the pointers are all --rcu and already visible to readers, as suggested by Eric Dumazet and Josh Triplett.

Hang on a moment. Do *none* of these cases need write memory barriers?

- Josh Triplet

#16 lore raw
Quoted 8 lines · Josh Triplett
Quoted 5 lines · Paul E. McKenney

Changes from v2:

o Switch from rcu_assign_pointer() to ACCESS_ONCE() given that
the pointers are all --rcu and already visible to readers, as suggested by Eric Dumazet and Josh Triplett.

Hang on a moment. Do *none* of these cases need write memory barriers?

Sigh. Some afternoons it doesn't pay to touch the keyboard.

Thank you for catching this. I will fix, but at this point, I am thinking in terms of 3.14 rather than 3.13 for this series.

Thanx, Paul

#17 lore raw
Quoted 8 lines · Paul E. McKenney
Quoted 7 lines · Josh Triplett
Quoted 5 lines · Paul E. McKenney

Changes from v2:

o Switch from rcu_assign_pointer() to ACCESS_ONCE() given that
the pointers are all --rcu and already visible to readers, as suggested by Eric Dumazet and Josh Triplett.

Quoted 6 lines
Quoted 1 line

Hang on a moment. Do *none* of these cases need write memory barriers?

Sigh. Some afternoons it doesn't pay to touch the keyboard.

Thank you for catching this. I will fix, but at this point, I am thinking in terms of 3.14 rather than 3.13 for this series.

Some of them looked safe. You could also replace --rcu with __rcu in the comments while at it.

Thanks,

Hannes

#18 lore raw
Quoted 11 lines · Hannes Frederic Sowa
Quoted 10 lines · Paul E. McKenney
Quoted 8 lines · Josh Triplett
Quoted 5 lines · Paul E. McKenney

Changes from v2:

o Switch from rcu_assign_pointer() to ACCESS_ONCE() given that
the pointers are all --rcu and already visible to readers, as suggested by Eric Dumazet and Josh Triplett.

Hang on a moment. Do *none* of these cases need write memory barriers?

Quoted 7 lines
Quoted 4 lines

Sigh. Some afternoons it doesn't pay to touch the keyboard.

Thank you for catching this. I will fix, but at this point, I am thinking in terms of 3.14 rather than 3.13 for this series.

Some of them looked safe. You could also replace --rcu with __rcu in the comments while at it.

Most of them deal with management, maybe a rtnl_assign_pointer with lockdep check for rtnl lock could help to not clean up the wrong bits.

I don't know if rtnl_assign_pointer is that a could name as it does not really explain why the barrier is not needed there. :/

Greetings,

Hannes

#20 lore raw
In reply to Hannes Frederic Sowa
Quoted 11 lines · Hannes Frederic Sowa
Quoted 10 lines · Paul E. McKenney
Quoted 8 lines · Josh Triplett
Quoted 5 lines · Paul E. McKenney

Changes from v2:

o Switch from rcu_assign_pointer() to ACCESS_ONCE() given that
the pointers are all --rcu and already visible to readers, as suggested by Eric Dumazet and Josh Triplett.

Hang on a moment. Do *none* of these cases need write memory barriers?

Quoted 7 lines
Quoted 4 lines

Sigh. Some afternoons it doesn't pay to touch the keyboard.

Thank you for catching this. I will fix, but at this point, I am thinking in terms of 3.14 rather than 3.13 for this series.

Some of them looked safe. You could also replace --rcu with __rcu in the comments while at it.

Sigh! Will do, thank you!

Thanx, Paul

#19 lore raw
In reply to Hannes Frederic Sowa
Quoted 17 lines · Hannes Frederic Sowa
Quoted 16 lines · Hannes Frederic Sowa
Quoted 14 lines · Paul E. McKenney
Quoted 8 lines · Josh Triplett
Quoted 5 lines · Paul E. McKenney

Changes from v2:

o Switch from rcu_assign_pointer() to ACCESS_ONCE() given that
the pointers are all --rcu and already visible to readers, as suggested by Eric Dumazet and Josh Triplett.

Hang on a moment. Do *none* of these cases need write memory barriers?

Sigh. Some afternoons it doesn't pay to touch the keyboard.

Thank you for catching this. I will fix, but at this point, I am thinking in terms of 3.14 rather than 3.13 for this series.

Quoted 8 lines
Quoted 2 lines

Some of them looked safe. You could also replace --rcu with __rcu in the comments while at it.

Most of them deal with management, maybe a rtnl_assign_pointer with lockdep check for rtnl lock could help to not clean up the wrong bits.

I don't know if rtnl_assign_pointer is that a could name as it does not really explain why the barrier is not needed there. :/

Beyond a certain point, I need to let people who know more about Linux's networking implementation handle this sort of thing.

Thanx, Paul