[PATCH v2 net-next 0/6] arp: Random clean up and RCU conversion for ioctl(SIOCGARP).

[PATCH v2 net-next 0/6] arp: Random clean up and RCU conversion for ioctl(SIOCGARP).

[Kuniyuki Iwashima]

arp_ioctl() holds rtnl_lock() regardless of cmd (SIOCDARP, SIOCSARP,
and SIOCGARP) to get net_device by __dev_get_by_name().

In the SIOCGARP path, arp_req_get() calls neigh_lookup(), which looks
up a neighbour entry under RCU.

This series cleans up ioctl() code a bit and extends the RCU section not
to take rtnl_lock() and instead use dev_get_by_name_rcu() for SIOCGARP.


Changes:
  v2:
    Patch 5: s/!IS_ERR/IS_ERR/ in arp_req_delete().

  v1: https://lore.kernel.org/netdev/20240422194755.4221-1-kuniyu@amazon.com/


Kuniyuki Iwashima (6):
  arp: Move ATF_COM setting in arp_req_set().
  arp: Validate netmask earlier for SIOCDARP and SIOCSARP in
    arp_ioctl().
  arp: Factorise ip_route_output() call in arp_req_set() and
    arp_req_delete().
  arp: Remove a nest in arp_req_get().
  arp: Get dev after calling arp_req_(delete|set|get)().
  arp: Convert ioctl(SIOCGARP) to RCU.

 net/ipv4/arp.c | 203 ++++++++++++++++++++++++++++++-------------------
 1 file changed, 123 insertions(+), 80 deletions(-)

-- 
2.30.2

[PATCH v2 net-next 1/6] arp: Move ATF_COM setting in arp_req_set().

[Kuniyuki Iwashima]

In arp_req_set(), if ATF_PERM is set in arpreq.arp_flags,
ATF_COM is set automatically.

The flag will be used later for neigh_update() only when
a neighbour entry is found.

Let's set ATF_COM just before calling neigh_update().

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
---
 net/ipv4/arp.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
index ab82ca104496..3093374165fa 100644
--- a/net/ipv4/arp.c
+++ b/net/ipv4/arp.c
@@ -1054,8 +1054,7 @@ static int arp_req_set(struct net *net, struct arpreq *r,
 		return arp_req_set_public(net, r, dev);
 
 	ip = ((struct sockaddr_in *)&r->arp_pa)->sin_addr.s_addr;
-	if (r->arp_flags & ATF_PERM)
-		r->arp_flags |= ATF_COM;
+
 	if (!dev) {
 		struct rtable *rt = ip_route_output(net, ip, 0, 0, 0,
 						    RT_SCOPE_LINK);
@@ -1092,8 +1091,12 @@ static int arp_req_set(struct net *net, struct arpreq *r,
 	err = PTR_ERR(neigh);
 	if (!IS_ERR(neigh)) {
 		unsigned int state = NUD_STALE;
-		if (r->arp_flags & ATF_PERM)
+
+		if (r->arp_flags & ATF_PERM) {
+			r->arp_flags |= ATF_COM;
 			state = NUD_PERMANENT;
+		}
+
 		err = neigh_update(neigh, (r->arp_flags & ATF_COM) ?
 				   r->arp_ha.sa_data : NULL, state,
 				   NEIGH_UPDATE_F_OVERRIDE |
-- 
2.30.2

[PATCH v2 net-next 2/6] arp: Validate netmask earlier for SIOCDARP and SIOCSARP in arp_ioctl().

[Kuniyuki Iwashima]

When ioctl(SIOCDARP/SIOCSARP) is issued with ATF_PUBL, r.arp_netmask
must be 0.0.0.0 or 255.255.255.255.

Currently, the netmask is validated in arp_req_delete_public() or
arp_req_set_public() under rtnl_lock().

We have ATF_NETMASK test in arp_ioctl() before holding rtnl_lock(),
so let's move the netmask validation there.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
---
 net/ipv4/arp.c | 27 +++++++++++++++------------
 1 file changed, 15 insertions(+), 12 deletions(-)

diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
index 3093374165fa..b20a5771d069 100644
--- a/net/ipv4/arp.c
+++ b/net/ipv4/arp.c
@@ -1023,11 +1023,8 @@ static int arp_req_set_proxy(struct net *net, struct net_device *dev, int on)
 static int arp_req_set_public(struct net *net, struct arpreq *r,
 		struct net_device *dev)
 {
-	__be32 ip = ((struct sockaddr_in *)&r->arp_pa)->sin_addr.s_addr;
 	__be32 mask = ((struct sockaddr_in *)&r->arp_netmask)->sin_addr.s_addr;
 
-	if (mask && mask != htonl(0xFFFFFFFF))
-		return -EINVAL;
 	if (!dev && (r->arp_flags & ATF_COM)) {
 		dev = dev_getbyhwaddr_rcu(net, r->arp_ha.sa_family,
 				      r->arp_ha.sa_data);
@@ -1035,6 +1032,8 @@ static int arp_req_set_public(struct net *net, struct arpreq *r,
 			return -ENODEV;
 	}
 	if (mask) {
+		__be32 ip = ((struct sockaddr_in *)&r->arp_pa)->sin_addr.s_addr;
+
 		if (!pneigh_lookup(&arp_tbl, net, &ip, dev, 1))
 			return -ENOBUFS;
 		return 0;
@@ -1171,14 +1170,13 @@ int arp_invalidate(struct net_device *dev, __be32 ip, bool force)
 static int arp_req_delete_public(struct net *net, struct arpreq *r,
 		struct net_device *dev)
 {
-	__be32 ip = ((struct sockaddr_in *) &r->arp_pa)->sin_addr.s_addr;
 	__be32 mask = ((struct sockaddr_in *)&r->arp_netmask)->sin_addr.s_addr;
 
-	if (mask == htonl(0xFFFFFFFF))
-		return pneigh_delete(&arp_tbl, net, &ip, dev);
+	if (mask) {
+		__be32 ip = ((struct sockaddr_in *)&r->arp_pa)->sin_addr.s_addr;
 
-	if (mask)
-		return -EINVAL;
+		return pneigh_delete(&arp_tbl, net, &ip, dev);
+	}
 
 	return arp_req_set_proxy(net, dev, 0);
 }
@@ -1211,9 +1209,10 @@ static int arp_req_delete(struct net *net, struct arpreq *r,
 
 int arp_ioctl(struct net *net, unsigned int cmd, void __user *arg)
 {
-	int err;
-	struct arpreq r;
 	struct net_device *dev = NULL;
+	struct arpreq r;
+	__be32 *netmask;
+	int err;
 
 	switch (cmd) {
 	case SIOCDARP:
@@ -1236,9 +1235,13 @@ int arp_ioctl(struct net *net, unsigned int cmd, void __user *arg)
 	if (!(r.arp_flags & ATF_PUBL) &&
 	    (r.arp_flags & (ATF_NETMASK | ATF_DONTPUB)))
 		return -EINVAL;
+
+	netmask = &((struct sockaddr_in *)&r.arp_netmask)->sin_addr.s_addr;
 	if (!(r.arp_flags & ATF_NETMASK))
-		((struct sockaddr_in *)&r.arp_netmask)->sin_addr.s_addr =
-							   htonl(0xFFFFFFFFUL);
+		*netmask = htonl(0xFFFFFFFFUL);
+	else if (*netmask && *netmask != htonl(0xFFFFFFFFUL))
+		return -EINVAL;
+
 	rtnl_lock();
 	if (r.arp_dev[0]) {
 		err = -ENODEV;
-- 
2.30.2

[PATCH v2 net-next 3/6] arp: Factorise ip_route_output() call in arp_req_set() and arp_req_delete().

[Kuniyuki Iwashima]

When ioctl(SIOCDARP/SIOCSARP) is issued for non-proxy entry (no ATF_COM)
without arpreq.arp_dev[] set, arp_req_set() and arp_req_delete() looks up
dev based on IPv4 address by ip_route_output().

Let's factorise the same code as arp_req_dev().

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
---
 net/ipv4/arp.c | 50 ++++++++++++++++++++++++++++++--------------------
 1 file changed, 30 insertions(+), 20 deletions(-)

diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
index b20a5771d069..ac3e15799c2f 100644
--- a/net/ipv4/arp.c
+++ b/net/ipv4/arp.c
@@ -1003,6 +1003,27 @@ static int arp_rcv(struct sk_buff *skb, struct net_device *dev,
  *	User level interface (ioctl)
  */
 
+static struct net_device *arp_req_dev(struct net *net, struct arpreq *r)
+{
+	struct net_device *dev;
+	struct rtable *rt;
+	__be32 ip;
+
+	ip = ((struct sockaddr_in *)&r->arp_pa)->sin_addr.s_addr;
+
+	rt = ip_route_output(net, ip, 0, 0, 0, RT_SCOPE_LINK);
+	if (IS_ERR(rt))
+		return ERR_CAST(rt);
+
+	dev = rt->dst.dev;
+	ip_rt_put(rt);
+
+	if (!dev)
+		return ERR_PTR(-EINVAL);
+
+	return dev;
+}
+
 /*
  *	Set (create) an ARP cache entry.
  */
@@ -1045,25 +1066,17 @@ static int arp_req_set_public(struct net *net, struct arpreq *r,
 static int arp_req_set(struct net *net, struct arpreq *r,
 		       struct net_device *dev)
 {
-	__be32 ip;
 	struct neighbour *neigh;
+	__be32 ip;
 	int err;
 
 	if (r->arp_flags & ATF_PUBL)
 		return arp_req_set_public(net, r, dev);
 
-	ip = ((struct sockaddr_in *)&r->arp_pa)->sin_addr.s_addr;
-
 	if (!dev) {
-		struct rtable *rt = ip_route_output(net, ip, 0, 0, 0,
-						    RT_SCOPE_LINK);
-
-		if (IS_ERR(rt))
-			return PTR_ERR(rt);
-		dev = rt->dst.dev;
-		ip_rt_put(rt);
-		if (!dev)
-			return -EINVAL;
+		dev = arp_req_dev(net, r);
+		if (IS_ERR(dev))
+			return PTR_ERR(dev);
 	}
 	switch (dev->type) {
 #if IS_ENABLED(CONFIG_FDDI)
@@ -1086,6 +1099,8 @@ static int arp_req_set(struct net *net, struct arpreq *r,
 		break;
 	}
 
+	ip = ((struct sockaddr_in *)&r->arp_pa)->sin_addr.s_addr;
+
 	neigh = __neigh_lookup_errno(&arp_tbl, &ip, dev);
 	err = PTR_ERR(neigh);
 	if (!IS_ERR(neigh)) {
@@ -1191,14 +1206,9 @@ static int arp_req_delete(struct net *net, struct arpreq *r,
 
 	ip = ((struct sockaddr_in *)&r->arp_pa)->sin_addr.s_addr;
 	if (!dev) {
-		struct rtable *rt = ip_route_output(net, ip, 0, 0, 0,
-						    RT_SCOPE_LINK);
-		if (IS_ERR(rt))
-			return PTR_ERR(rt);
-		dev = rt->dst.dev;
-		ip_rt_put(rt);
-		if (!dev)
-			return -EINVAL;
+		dev = arp_req_dev(net, r);
+		if (IS_ERR(dev))
+			return PTR_ERR(dev);
 	}
 	return arp_invalidate(dev, ip, true);
 }
-- 
2.30.2

[PATCH v2 net-next 4/6] arp: Remove a nest in arp_req_get().

[Kuniyuki Iwashima]

This is a prep patch to make the following changes tidy.

No functional change intended.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
---
 net/ipv4/arp.c | 31 ++++++++++++++++++-------------
 1 file changed, 18 insertions(+), 13 deletions(-)

diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
index ac3e15799c2f..60f633b24ec8 100644
--- a/net/ipv4/arp.c
+++ b/net/ipv4/arp.c
@@ -1138,23 +1138,28 @@ static int arp_req_get(struct arpreq *r, struct net_device *dev)
 {
 	__be32 ip = ((struct sockaddr_in *) &r->arp_pa)->sin_addr.s_addr;
 	struct neighbour *neigh;
-	int err = -ENXIO;
 
 	neigh = neigh_lookup(&arp_tbl, &ip, dev);
-	if (neigh) {
-		if (!(READ_ONCE(neigh->nud_state) & NUD_NOARP)) {
-			read_lock_bh(&neigh->lock);
-			memcpy(r->arp_ha.sa_data, neigh->ha,
-			       min(dev->addr_len, sizeof(r->arp_ha.sa_data_min)));
-			r->arp_flags = arp_state_to_flags(neigh);
-			read_unlock_bh(&neigh->lock);
-			r->arp_ha.sa_family = dev->type;
-			strscpy(r->arp_dev, dev->name, sizeof(r->arp_dev));
-			err = 0;
-		}
+	if (!neigh)
+		return -ENXIO;
+
+	if (READ_ONCE(neigh->nud_state) & NUD_NOARP) {
 		neigh_release(neigh);
+		return -ENXIO;
 	}
-	return err;
+
+	read_lock_bh(&neigh->lock);
+	memcpy(r->arp_ha.sa_data, neigh->ha,
+	       min(dev->addr_len, sizeof(r->arp_ha.sa_data_min)));
+	r->arp_flags = arp_state_to_flags(neigh);
+	read_unlock_bh(&neigh->lock);
+
+	neigh_release(neigh);
+
+	r->arp_ha.sa_family = dev->type;
+	strscpy(r->arp_dev, dev->name, sizeof(r->arp_dev));
+
+	return 0;
 }
 
 int arp_invalidate(struct net_device *dev, __be32 ip, bool force)
-- 
2.30.2

[PATCH v2 net-next 5/6] arp: Get dev after calling arp_req_(delete|set|get)().

[Kuniyuki Iwashima]

arp_ioctl() holds rtnl_lock() first regardless of cmd (SIOCDARP,
SIOCSARP, and SIOCGARP) to get net_device by __dev_get_by_name().

In the SIOCGARP path, arp_req_get() calls neigh_lookup(), which
looks up a neighbour entry under RCU.

We will extend the RCU section not to take rtnl_lock() and instead
use dev_get_by_name_rcu() for SIOCGARP.

As a preparation, let's move __dev_get_by_name() into another
function and call it from arp_req_delete(), arp_req_set(), and
arp_req_get().

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
---
 net/ipv4/arp.c | 86 +++++++++++++++++++++++++++++---------------------
 1 file changed, 50 insertions(+), 36 deletions(-)

diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
index 60f633b24ec8..5034920be85a 100644
--- a/net/ipv4/arp.c
+++ b/net/ipv4/arp.c
@@ -1003,12 +1003,36 @@ static int arp_rcv(struct sk_buff *skb, struct net_device *dev,
  *	User level interface (ioctl)
  */
 
+static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r)
+{
+	struct net_device *dev;
+
+	dev = __dev_get_by_name(net, r->arp_dev);
+	if (!dev)
+		return ERR_PTR(-ENODEV);
+
+	/* Mmmm... It is wrong... ARPHRD_NETROM == 0 */
+	if (!r->arp_ha.sa_family)
+		r->arp_ha.sa_family = dev->type;
+
+	if ((r->arp_flags & ATF_COM) && r->arp_ha.sa_family != dev->type)
+		return ERR_PTR(-EINVAL);
+
+	return dev;
+}
+
 static struct net_device *arp_req_dev(struct net *net, struct arpreq *r)
 {
 	struct net_device *dev;
 	struct rtable *rt;
 	__be32 ip;
 
+	if (r->arp_dev[0])
+		return arp_req_dev_by_name(net, r);
+
+	if (r->arp_flags & ATF_PUBL)
+		return NULL;
+
 	ip = ((struct sockaddr_in *)&r->arp_pa)->sin_addr.s_addr;
 
 	rt = ip_route_output(net, ip, 0, 0, 0, RT_SCOPE_LINK);
@@ -1063,21 +1087,20 @@ static int arp_req_set_public(struct net *net, struct arpreq *r,
 	return arp_req_set_proxy(net, dev, 1);
 }
 
-static int arp_req_set(struct net *net, struct arpreq *r,
-		       struct net_device *dev)
+static int arp_req_set(struct net *net, struct arpreq *r)
 {
 	struct neighbour *neigh;
+	struct net_device *dev;
 	__be32 ip;
 	int err;
 
+	dev = arp_req_dev(net, r);
+	if (IS_ERR(dev))
+		return PTR_ERR(dev);
+
 	if (r->arp_flags & ATF_PUBL)
 		return arp_req_set_public(net, r, dev);
 
-	if (!dev) {
-		dev = arp_req_dev(net, r);
-		if (IS_ERR(dev))
-			return PTR_ERR(dev);
-	}
 	switch (dev->type) {
 #if IS_ENABLED(CONFIG_FDDI)
 	case ARPHRD_FDDI:
@@ -1134,10 +1157,18 @@ static unsigned int arp_state_to_flags(struct neighbour *neigh)
  *	Get an ARP cache entry.
  */
 
-static int arp_req_get(struct arpreq *r, struct net_device *dev)
+static int arp_req_get(struct net *net, struct arpreq *r)
 {
 	__be32 ip = ((struct sockaddr_in *) &r->arp_pa)->sin_addr.s_addr;
 	struct neighbour *neigh;
+	struct net_device *dev;
+
+	if (!r->arp_dev[0])
+		return -ENODEV;
+
+	dev = arp_req_dev_by_name(net, r);
+	if (IS_ERR(dev))
+		return PTR_ERR(dev);
 
 	neigh = neigh_lookup(&arp_tbl, &ip, dev);
 	if (!neigh)
@@ -1201,20 +1232,20 @@ static int arp_req_delete_public(struct net *net, struct arpreq *r,
 	return arp_req_set_proxy(net, dev, 0);
 }
 
-static int arp_req_delete(struct net *net, struct arpreq *r,
-			  struct net_device *dev)
+static int arp_req_delete(struct net *net, struct arpreq *r)
 {
+	struct net_device *dev;
 	__be32 ip;
 
+	dev = arp_req_dev(net, r);
+	if (IS_ERR(dev))
+		return PTR_ERR(dev);
+
 	if (r->arp_flags & ATF_PUBL)
 		return arp_req_delete_public(net, r, dev);
 
 	ip = ((struct sockaddr_in *)&r->arp_pa)->sin_addr.s_addr;
-	if (!dev) {
-		dev = arp_req_dev(net, r);
-		if (IS_ERR(dev))
-			return PTR_ERR(dev);
-	}
+
 	return arp_invalidate(dev, ip, true);
 }
 
@@ -1224,7 +1255,6 @@ static int arp_req_delete(struct net *net, struct arpreq *r,
 
 int arp_ioctl(struct net *net, unsigned int cmd, void __user *arg)
 {
-	struct net_device *dev = NULL;
 	struct arpreq r;
 	__be32 *netmask;
 	int err;
@@ -1258,35 +1288,19 @@ int arp_ioctl(struct net *net, unsigned int cmd, void __user *arg)
 		return -EINVAL;
 
 	rtnl_lock();
-	if (r.arp_dev[0]) {
-		err = -ENODEV;
-		dev = __dev_get_by_name(net, r.arp_dev);
-		if (!dev)
-			goto out;
-
-		/* Mmmm... It is wrong... ARPHRD_NETROM==0 */
-		if (!r.arp_ha.sa_family)
-			r.arp_ha.sa_family = dev->type;
-		err = -EINVAL;
-		if ((r.arp_flags & ATF_COM) && r.arp_ha.sa_family != dev->type)
-			goto out;
-	} else if (cmd == SIOCGARP) {
-		err = -ENODEV;
-		goto out;
-	}
 
 	switch (cmd) {
 	case SIOCDARP:
-		err = arp_req_delete(net, &r, dev);
+		err = arp_req_delete(net, &r);
 		break;
 	case SIOCSARP:
-		err = arp_req_set(net, &r, dev);
+		err = arp_req_set(net, &r);
 		break;
 	case SIOCGARP:
-		err = arp_req_get(&r, dev);
+		err = arp_req_get(net, &r);
 		break;
 	}
-out:
+
 	rtnl_unlock();
 	if (cmd == SIOCGARP && !err && copy_to_user(arg, &r, sizeof(r)))
 		err = -EFAULT;
-- 
2.30.2

[PATCH v2 net-next 6/6] arp: Convert ioctl(SIOCGARP) to RCU.

[Kuniyuki Iwashima]

ioctl(SIOCGARP) holds rtnl_lock() for __dev_get_by_name() and
later calls neigh_lookup(), which calls rcu_read_lock().

Let's replace __dev_get_by_name() with dev_get_by_name_rcu() to
avoid locking rtnl_lock().

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
---
 net/ipv4/arp.c | 26 +++++++++++++++++---------
 1 file changed, 17 insertions(+), 9 deletions(-)

diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
index 5034920be85a..9430b64558cd 100644
--- a/net/ipv4/arp.c
+++ b/net/ipv4/arp.c
@@ -1003,11 +1003,15 @@ static int arp_rcv(struct sk_buff *skb, struct net_device *dev,
  *	User level interface (ioctl)
  */
 
-static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r)
+static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r,
+					      bool getarp)
 {
 	struct net_device *dev;
 
-	dev = __dev_get_by_name(net, r->arp_dev);
+	if (getarp)
+		dev = dev_get_by_name_rcu(net, r->arp_dev);
+	else
+		dev = __dev_get_by_name(net, r->arp_dev);
 	if (!dev)
 		return ERR_PTR(-ENODEV);
 
@@ -1028,7 +1032,7 @@ static struct net_device *arp_req_dev(struct net *net, struct arpreq *r)
 	__be32 ip;
 
 	if (r->arp_dev[0])
-		return arp_req_dev_by_name(net, r);
+		return arp_req_dev_by_name(net, r, false);
 
 	if (r->arp_flags & ATF_PUBL)
 		return NULL;
@@ -1166,7 +1170,7 @@ static int arp_req_get(struct net *net, struct arpreq *r)
 	if (!r->arp_dev[0])
 		return -ENODEV;
 
-	dev = arp_req_dev_by_name(net, r);
+	dev = arp_req_dev_by_name(net, r, true);
 	if (IS_ERR(dev))
 		return PTR_ERR(dev);
 
@@ -1287,23 +1291,27 @@ int arp_ioctl(struct net *net, unsigned int cmd, void __user *arg)
 	else if (*netmask && *netmask != htonl(0xFFFFFFFFUL))
 		return -EINVAL;
 
-	rtnl_lock();
-
 	switch (cmd) {
 	case SIOCDARP:
+		rtnl_lock();
 		err = arp_req_delete(net, &r);
+		rtnl_unlock();
 		break;
 	case SIOCSARP:
+		rtnl_lock();
 		err = arp_req_set(net, &r);
+		rtnl_unlock();
 		break;
 	case SIOCGARP:
+		rcu_read_lock();
 		err = arp_req_get(net, &r);
+		rcu_read_unlock();
+
+		if (!err && copy_to_user(arg, &r, sizeof(r)))
+			err = -EFAULT;
 		break;
 	}
 
-	rtnl_unlock();
-	if (cmd == SIOCGARP && !err && copy_to_user(arg, &r, sizeof(r)))
-		err = -EFAULT;
 	return err;
 }
 
-- 
2.30.2

Re: [PATCH v2 net-next 6/6] arp: Convert ioctl(SIOCGARP) to RCU.

[Eric Dumazet]

On Thu, Apr 25, 2024 at 7:02 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
>
> ioctl(SIOCGARP) holds rtnl_lock() for __dev_get_by_name() and
> later calls neigh_lookup(), which calls rcu_read_lock().
>
> Let's replace __dev_get_by_name() with dev_get_by_name_rcu() to
> avoid locking rtnl_lock().
>
> Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
> ---
>  net/ipv4/arp.c | 26 +++++++++++++++++---------
>  1 file changed, 17 insertions(+), 9 deletions(-)
>
> diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
> index 5034920be85a..9430b64558cd 100644
> --- a/net/ipv4/arp.c
> +++ b/net/ipv4/arp.c
> @@ -1003,11 +1003,15 @@ static int arp_rcv(struct sk_buff *skb, struct net_device *dev,
>   *     User level interface (ioctl)
>   */
>
> -static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r)
> +static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r,
> +                                             bool getarp)
>  {
>         struct net_device *dev;
>
> -       dev = __dev_get_by_name(net, r->arp_dev);
> +       if (getarp)
> +               dev = dev_get_by_name_rcu(net, r->arp_dev);
> +       else
> +               dev = __dev_get_by_name(net, r->arp_dev);
>         if (!dev)
>                 return ERR_PTR(-ENODEV);
>
> @@ -1028,7 +1032,7 @@ static struct net_device *arp_req_dev(struct net *net, struct arpreq *r)
>         __be32 ip;
>
>         if (r->arp_dev[0])
> -               return arp_req_dev_by_name(net, r);
> +               return arp_req_dev_by_name(net, r, false);
>
>         if (r->arp_flags & ATF_PUBL)
>                 return NULL;
> @@ -1166,7 +1170,7 @@ static int arp_req_get(struct net *net, struct arpreq *r)
>         if (!r->arp_dev[0])
>                 return -ENODEV;
>
> -       dev = arp_req_dev_by_name(net, r);
> +       dev = arp_req_dev_by_name(net, r, true);
>         if (IS_ERR(dev))
>                 return PTR_ERR(dev);
>
> @@ -1287,23 +1291,27 @@ int arp_ioctl(struct net *net, unsigned int cmd, void __user *arg)
>         else if (*netmask && *netmask != htonl(0xFFFFFFFFUL))
>                 return -EINVAL;
>
> -       rtnl_lock();
> -
>         switch (cmd) {
>         case SIOCDARP:
> +               rtnl_lock();
>                 err = arp_req_delete(net, &r);
> +               rtnl_unlock();
>                 break;
>         case SIOCSARP:
> +               rtnl_lock();
>                 err = arp_req_set(net, &r);
> +               rtnl_unlock();
>                 break;
>         case SIOCGARP:
> +               rcu_read_lock();
>                 err = arp_req_get(net, &r);
> +               rcu_read_unlock();


Note that arp_req_get() uses :

strscpy(r->arp_dev, dev->name, sizeof(r->arp_dev));

This currently depends on RTNL or devnet_rename_sem

Perhaps we should add a helper and use a seqlock to safely copy
dev->name into a temporary variable.

netdev_get_name() can not be called from rcu_read_lock() at this
moment unfortunately.

Re: [PATCH v2 net-next 6/6] arp: Convert ioctl(SIOCGARP) to RCU.

[Kuniyuki Iwashima]

From: Eric Dumazet <edumazet@google.com>
Date: Thu, 25 Apr 2024 19:12:56 +0200
> On Thu, Apr 25, 2024 at 7:02 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> >
> > ioctl(SIOCGARP) holds rtnl_lock() for __dev_get_by_name() and
> > later calls neigh_lookup(), which calls rcu_read_lock().
> >
> > Let's replace __dev_get_by_name() with dev_get_by_name_rcu() to
> > avoid locking rtnl_lock().
> >
> > Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
> > ---
> >  net/ipv4/arp.c | 26 +++++++++++++++++---------
> >  1 file changed, 17 insertions(+), 9 deletions(-)
> >
> > diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
> > index 5034920be85a..9430b64558cd 100644
> > --- a/net/ipv4/arp.c
> > +++ b/net/ipv4/arp.c
> > @@ -1003,11 +1003,15 @@ static int arp_rcv(struct sk_buff *skb, struct net_device *dev,
> >   *     User level interface (ioctl)
> >   */
> >
> > -static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r)
> > +static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r,
> > +                                             bool getarp)
> >  {
> >         struct net_device *dev;
> >
> > -       dev = __dev_get_by_name(net, r->arp_dev);
> > +       if (getarp)
> > +               dev = dev_get_by_name_rcu(net, r->arp_dev);
> > +       else
> > +               dev = __dev_get_by_name(net, r->arp_dev);
> >         if (!dev)
> >                 return ERR_PTR(-ENODEV);
> >
> > @@ -1028,7 +1032,7 @@ static struct net_device *arp_req_dev(struct net *net, struct arpreq *r)
> >         __be32 ip;
> >
> >         if (r->arp_dev[0])
> > -               return arp_req_dev_by_name(net, r);
> > +               return arp_req_dev_by_name(net, r, false);
> >
> >         if (r->arp_flags & ATF_PUBL)
> >                 return NULL;
> > @@ -1166,7 +1170,7 @@ static int arp_req_get(struct net *net, struct arpreq *r)
> >         if (!r->arp_dev[0])
> >                 return -ENODEV;
> >
> > -       dev = arp_req_dev_by_name(net, r);
> > +       dev = arp_req_dev_by_name(net, r, true);
> >         if (IS_ERR(dev))
> >                 return PTR_ERR(dev);
> >
> > @@ -1287,23 +1291,27 @@ int arp_ioctl(struct net *net, unsigned int cmd, void __user *arg)
> >         else if (*netmask && *netmask != htonl(0xFFFFFFFFUL))
> >                 return -EINVAL;
> >
> > -       rtnl_lock();
> > -
> >         switch (cmd) {
> >         case SIOCDARP:
> > +               rtnl_lock();
> >                 err = arp_req_delete(net, &r);
> > +               rtnl_unlock();
> >                 break;
> >         case SIOCSARP:
> > +               rtnl_lock();
> >                 err = arp_req_set(net, &r);
> > +               rtnl_unlock();
> >                 break;
> >         case SIOCGARP:
> > +               rcu_read_lock();
> >                 err = arp_req_get(net, &r);
> > +               rcu_read_unlock();
> 
> 
> Note that arp_req_get() uses :
> 
> strscpy(r->arp_dev, dev->name, sizeof(r->arp_dev));
> 
> This currently depends on RTNL or devnet_rename_sem

Ah, I missed this point, thanks for catching!


> 
> Perhaps we should add a helper and use a seqlock to safely copy
> dev->name into a temporary variable.

So it's preferable to add seqlock around memcpy() in dev_change_name()
and use a helper in arp_req_get() rather than adding devnet_rename_sem
locking around memcpy() in arp_req_get() ?

Re: [PATCH v2 net-next 6/6] arp: Convert ioctl(SIOCGARP) to RCU.

[Eric Dumazet]

On Thu, Apr 25, 2024 at 7:52 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
>
> From: Eric Dumazet <edumazet@google.com>
> Date: Thu, 25 Apr 2024 19:12:56 +0200
> > On Thu, Apr 25, 2024 at 7:02 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> > >
> > > ioctl(SIOCGARP) holds rtnl_lock() for __dev_get_by_name() and
> > > later calls neigh_lookup(), which calls rcu_read_lock().
> > >
> > > Let's replace __dev_get_by_name() with dev_get_by_name_rcu() to
> > > avoid locking rtnl_lock().
> > >
> > > Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
> > > ---
> > >  net/ipv4/arp.c | 26 +++++++++++++++++---------
> > >  1 file changed, 17 insertions(+), 9 deletions(-)
> > >
> > > diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
> > > index 5034920be85a..9430b64558cd 100644
> > > --- a/net/ipv4/arp.c
> > > +++ b/net/ipv4/arp.c
> > > @@ -1003,11 +1003,15 @@ static int arp_rcv(struct sk_buff *skb, struct net_device *dev,
> > >   *     User level interface (ioctl)
> > >   */
> > >
> > > -static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r)
> > > +static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r,
> > > +                                             bool getarp)
> > >  {
> > >         struct net_device *dev;
> > >
> > > -       dev = __dev_get_by_name(net, r->arp_dev);
> > > +       if (getarp)
> > > +               dev = dev_get_by_name_rcu(net, r->arp_dev);
> > > +       else
> > > +               dev = __dev_get_by_name(net, r->arp_dev);
> > >         if (!dev)
> > >                 return ERR_PTR(-ENODEV);
> > >
> > > @@ -1028,7 +1032,7 @@ static struct net_device *arp_req_dev(struct net *net, struct arpreq *r)
> > >         __be32 ip;
> > >
> > >         if (r->arp_dev[0])
> > > -               return arp_req_dev_by_name(net, r);
> > > +               return arp_req_dev_by_name(net, r, false);
> > >
> > >         if (r->arp_flags & ATF_PUBL)
> > >                 return NULL;
> > > @@ -1166,7 +1170,7 @@ static int arp_req_get(struct net *net, struct arpreq *r)
> > >         if (!r->arp_dev[0])
> > >                 return -ENODEV;
> > >
> > > -       dev = arp_req_dev_by_name(net, r);
> > > +       dev = arp_req_dev_by_name(net, r, true);
> > >         if (IS_ERR(dev))
> > >                 return PTR_ERR(dev);
> > >
> > > @@ -1287,23 +1291,27 @@ int arp_ioctl(struct net *net, unsigned int cmd, void __user *arg)
> > >         else if (*netmask && *netmask != htonl(0xFFFFFFFFUL))
> > >                 return -EINVAL;
> > >
> > > -       rtnl_lock();
> > > -
> > >         switch (cmd) {
> > >         case SIOCDARP:
> > > +               rtnl_lock();
> > >                 err = arp_req_delete(net, &r);
> > > +               rtnl_unlock();
> > >                 break;
> > >         case SIOCSARP:
> > > +               rtnl_lock();
> > >                 err = arp_req_set(net, &r);
> > > +               rtnl_unlock();
> > >                 break;
> > >         case SIOCGARP:
> > > +               rcu_read_lock();
> > >                 err = arp_req_get(net, &r);
> > > +               rcu_read_unlock();
> >
> >
> > Note that arp_req_get() uses :
> >
> > strscpy(r->arp_dev, dev->name, sizeof(r->arp_dev));
> >
> > This currently depends on RTNL or devnet_rename_sem
>
> Ah, I missed this point, thanks for catching!
>
>
> >
> > Perhaps we should add a helper and use a seqlock to safely copy
> > dev->name into a temporary variable.
>
> So it's preferable to add seqlock around memcpy() in dev_change_name()
> and use a helper in arp_req_get() rather than adding devnet_rename_sem
> locking around memcpy() in arp_req_get() ?

Under rcu_read_lock(), we can not sleep.

devnet_rename_sem is a semaphore... down_read() might sleep.

So if you plan using current netdev_get_name(), you must call it
outside of rcu_read_lock() section.

Re: [PATCH v2 net-next 6/6] arp: Convert ioctl(SIOCGARP) to RCU.

[Kuniyuki Iwashima]

From: Eric Dumazet <edumazet@google.com>
Date: Thu, 25 Apr 2024 22:35:38 +0200
> On Thu, Apr 25, 2024 at 7:52 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> >
> > From: Eric Dumazet <edumazet@google.com>
> > Date: Thu, 25 Apr 2024 19:12:56 +0200
> > > On Thu, Apr 25, 2024 at 7:02 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> > > >
> > > > ioctl(SIOCGARP) holds rtnl_lock() for __dev_get_by_name() and
> > > > later calls neigh_lookup(), which calls rcu_read_lock().
> > > >
> > > > Let's replace __dev_get_by_name() with dev_get_by_name_rcu() to
> > > > avoid locking rtnl_lock().
> > > >
> > > > Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
> > > > ---
> > > >  net/ipv4/arp.c | 26 +++++++++++++++++---------
> > > >  1 file changed, 17 insertions(+), 9 deletions(-)
> > > >
> > > > diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
> > > > index 5034920be85a..9430b64558cd 100644
> > > > --- a/net/ipv4/arp.c
> > > > +++ b/net/ipv4/arp.c
> > > > @@ -1003,11 +1003,15 @@ static int arp_rcv(struct sk_buff *skb, struct net_device *dev,
> > > >   *     User level interface (ioctl)
> > > >   */
> > > >
> > > > -static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r)
> > > > +static struct net_device *arp_req_dev_by_name(struct net *net, struct arpreq *r,
> > > > +                                             bool getarp)
> > > >  {
> > > >         struct net_device *dev;
> > > >
> > > > -       dev = __dev_get_by_name(net, r->arp_dev);
> > > > +       if (getarp)
> > > > +               dev = dev_get_by_name_rcu(net, r->arp_dev);
> > > > +       else
> > > > +               dev = __dev_get_by_name(net, r->arp_dev);
> > > >         if (!dev)
> > > >                 return ERR_PTR(-ENODEV);
> > > >
> > > > @@ -1028,7 +1032,7 @@ static struct net_device *arp_req_dev(struct net *net, struct arpreq *r)
> > > >         __be32 ip;
> > > >
> > > >         if (r->arp_dev[0])
> > > > -               return arp_req_dev_by_name(net, r);
> > > > +               return arp_req_dev_by_name(net, r, false);
> > > >
> > > >         if (r->arp_flags & ATF_PUBL)
> > > >                 return NULL;
> > > > @@ -1166,7 +1170,7 @@ static int arp_req_get(struct net *net, struct arpreq *r)
> > > >         if (!r->arp_dev[0])
> > > >                 return -ENODEV;
> > > >
> > > > -       dev = arp_req_dev_by_name(net, r);
> > > > +       dev = arp_req_dev_by_name(net, r, true);
> > > >         if (IS_ERR(dev))
> > > >                 return PTR_ERR(dev);
> > > >
> > > > @@ -1287,23 +1291,27 @@ int arp_ioctl(struct net *net, unsigned int cmd, void __user *arg)
> > > >         else if (*netmask && *netmask != htonl(0xFFFFFFFFUL))
> > > >                 return -EINVAL;
> > > >
> > > > -       rtnl_lock();
> > > > -
> > > >         switch (cmd) {
> > > >         case SIOCDARP:
> > > > +               rtnl_lock();
> > > >                 err = arp_req_delete(net, &r);
> > > > +               rtnl_unlock();
> > > >                 break;
> > > >         case SIOCSARP:
> > > > +               rtnl_lock();
> > > >                 err = arp_req_set(net, &r);
> > > > +               rtnl_unlock();
> > > >                 break;
> > > >         case SIOCGARP:
> > > > +               rcu_read_lock();
> > > >                 err = arp_req_get(net, &r);
> > > > +               rcu_read_unlock();
> > >
> > >
> > > Note that arp_req_get() uses :
> > >
> > > strscpy(r->arp_dev, dev->name, sizeof(r->arp_dev));
> > >
> > > This currently depends on RTNL or devnet_rename_sem
> >
> > Ah, I missed this point, thanks for catching!
> >
> >
> > >
> > > Perhaps we should add a helper and use a seqlock to safely copy
> > > dev->name into a temporary variable.
> >
> > So it's preferable to add seqlock around memcpy() in dev_change_name()
> > and use a helper in arp_req_get() rather than adding devnet_rename_sem
> > locking around memcpy() in arp_req_get() ?
> 
> Under rcu_read_lock(), we can not sleep.
> 
> devnet_rename_sem is a semaphore... down_read() might sleep.

yes... -ENOCOFFEE :)


> 
> So if you plan using current netdev_get_name(), you must call it
> outside of rcu_read_lock() section.
>

Will add seqlock helper in v3.

Thanks!