Skip to contentSkip to conversation

um: vector: fix NULL pointer derefs in queue-less transports

PATCH In review Michael Bommarito 3 participants 3 messages Apr 10 – Apr 22

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 um: vector: fix NULL pointer derefs in queue-less transports

TAP transport sets neither VECTOR_RX nor VECTOR_TX, so vector_net_open() never allocates rx_queue or tx_queue. HYBRID sets VECTOR_RX but not VECTOR_TX, so tx_queue is NULL there too.

vector_reset_stats(), vector_poll(), vector_get_ethtool_stats(), and vector_get_ringparam() unconditionally deref these queue pointers, causing a NULL pointer crash on SMP or with any lock debugging option.

Guard all queue pointer accesses with NULL checks.

Fixes 49da7e64f33e ("High Performance UML Vector Network Driver")
Cc Anton Ivanov <[email protected]>
Assisted-by Claude:claude-opus-4-6
Assisted-by Codex:gpt-5-4

---
Found while enabling KCOV and lockdep on UML for a network-stack test lab. Tested boot with SMP=y + PROVE_LOCKING + DEBUG_SPINLOCK + DEBUG_LOCK_ALLOC + LOCKDEP + KCOV, all with vec0:transport=tap.

Without the fix, the same config panics at addr 0x18 (SMP, no debug), 0x1c (DEBUG_SPINLOCK), or 0x30 (lockdep) -- all offsets into a NULL vector_queue pointer.

 arch/um/drivers/vector_kern.c | 48 +++++++++++++++++------------------
 1 file changed, 24 insertions(+), 24 deletions(-)
arch/um/drivers/vector_kern.c +24 -24
index 2cc90055499a5..6134c376e57be 100644
@@ -105,25 +105,18 @@ static const struct {
 
 static void vector_reset_stats(struct vector_private *vp)
 {
-	/* We reuse the existing queue locks for stats */
-
-	/* RX stats are modified with RX head_lock held
-	 * in vector_poll.
-	 */
-
-	spin_lock(&vp->rx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_lock(&vp->rx_queue->head_lock);
 	vp->estats.rx_queue_max = 0;
 	vp->estats.rx_queue_running_average = 0;
 	vp->estats.rx_encaps_errors = 0;
 	vp->estats.sg_ok = 0;
 	vp->estats.sg_linearized = 0;
-	spin_unlock(&vp->rx_queue->head_lock);
-
-	/* TX stats are modified with TX head_lock held
-	 * in vector_send.
-	 */
+	if (vp->rx_queue)
+		spin_unlock(&vp->rx_queue->head_lock);
 
-	spin_lock(&vp->tx_queue->head_lock);
+	if (vp->tx_queue)
+		spin_lock(&vp->tx_queue->head_lock);
 	vp->estats.tx_timeout_count = 0;
 	vp->estats.tx_restart_queue = 0;
 	vp->estats.tx_kicks = 0;
@@ -131,7 +124,8 @@ static void vector_reset_stats(struct vector_private *vp)
 	vp->estats.tx_flow_control_xoff = 0;
 	vp->estats.tx_queue_max = 0;
 	vp->estats.tx_queue_running_average = 0;
-	spin_unlock(&vp->tx_queue->head_lock);
+	if (vp->tx_queue)
+		spin_unlock(&vp->tx_queue->head_lock);
 }
 
 static int get_mtu(struct arglist *def)
@@ -1163,7 +1157,8 @@ static int vector_poll(struct napi_struct *napi, int budget)
 
 	if ((vp->options & VECTOR_TX) != 0)
 		tx_enqueued = (vector_send(vp->tx_queue) > 0);
-	spin_lock(&vp->rx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_lock(&vp->rx_queue->head_lock);
 	if ((vp->options & VECTOR_RX) > 0)
 		err = vector_mmsg_rx(vp, budget);
 	else {
@@ -1171,7 +1166,8 @@ static int vector_poll(struct napi_struct *napi, int budget)
 		if (err > 0)
 			err = 1;
 	}
-	spin_unlock(&vp->rx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_unlock(&vp->rx_queue->head_lock);
 	if (err > 0)
 		work_done += err;
 
@@ -1421,10 +1417,10 @@ static void vector_get_ringparam(struct net_device *netdev,
 {
 	struct vector_private *vp = netdev_priv(netdev);
 
-	ring->rx_max_pending = vp->rx_queue->max_depth;
-	ring->tx_max_pending = vp->tx_queue->max_depth;
-	ring->rx_pending = vp->rx_queue->max_depth;
-	ring->tx_pending = vp->tx_queue->max_depth;
+	ring->rx_max_pending = vp->rx_queue ? vp->rx_queue->max_depth : 0;
+	ring->tx_max_pending = vp->tx_queue ? vp->tx_queue->max_depth : 0;
+	ring->rx_pending = ring->rx_max_pending;
+	ring->tx_pending = ring->tx_max_pending;
 }
 
 static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
@@ -1466,11 +1462,15 @@ static void vector_get_ethtool_stats(struct net_device *dev,
 	 * to date.
 	 */
 
-	spin_lock(&vp->tx_queue->head_lock);
-	spin_lock(&vp->rx_queue->head_lock);
+	if (vp->tx_queue)
+		spin_lock(&vp->tx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_lock(&vp->rx_queue->head_lock);
 	memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
-	spin_unlock(&vp->rx_queue->head_lock);
-	spin_unlock(&vp->tx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_unlock(&vp->rx_queue->head_lock);
+	if (vp->tx_queue)
+		spin_unlock(&vp->tx_queue->head_lock);
 }
 
 static int vector_get_coalesce(struct net_device *netdev,
Signature · 1 line
2.53.0
maintainer Acked-by on v1 1/1
#1 lore raw
Quoted 119 lines · Michael Bommarito

TAP transport sets neither VECTOR_RX nor VECTOR_TX, so vector_net_open() never allocates rx_queue or tx_queue. HYBRID sets VECTOR_RX but not VECTOR_TX, so tx_queue is NULL there too.

vector_reset_stats(), vector_poll(), vector_get_ethtool_stats(), and vector_get_ringparam() unconditionally deref these queue pointers, causing a NULL pointer crash on SMP or with any lock debugging option.

Guard all queue pointer accesses with NULL checks.

Fixes 49da7e64f33e ("High Performance UML Vector Network Driver")
Cc Anton Ivanov <[email protected]>
Assisted-by Claude:claude-opus-4-6
Assisted-by Codex:gpt-5-4

---
Found while enabling KCOV and lockdep on UML for a network-stack test lab. Tested boot with SMP=y + PROVE_LOCKING + DEBUG_SPINLOCK + DEBUG_LOCK_ALLOC + LOCKDEP + KCOV, all with vec0:transport=tap.

Without the fix, the same config panics at addr 0x18 (SMP, no debug), 0x1c (DEBUG_SPINLOCK), or 0x30 (lockdep) -- all offsets into a NULL vector_queue pointer.

  arch/um/drivers/vector_kern.c | 48 +++++++++++++++++------------------
  1 file changed, 24 insertions(+), 24 deletions(-)
arch/um/drivers/vector_kern.c +20 -22
index 2cc90055499a5..6134c376e57be 100644
@@ -105,25 +105,18 @@ static const struct {
 
  static void vector_reset_stats(struct vector_private *vp)
  {
-	/* We reuse the existing queue locks for stats */
-
-	/* RX stats are modified with RX head_lock held
-	 * in vector_poll.
-	 */
-
-	spin_lock(&vp->rx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_lock(&vp->rx_queue->head_lock);
  	vp->estats.rx_queue_max = 0;
  	vp->estats.rx_queue_running_average = 0;
  	vp->estats.rx_encaps_errors = 0;
  	vp->estats.sg_ok = 0;
  	vp->estats.sg_linearized = 0;
-	spin_unlock(&vp->rx_queue->head_lock);
-
-	/* TX stats are modified with TX head_lock held
-	 * in vector_send.
-	 */
+	if (vp->rx_queue)
+		spin_unlock(&vp->rx_queue->head_lock);
 
-	spin_lock(&vp->tx_queue->head_lock);
+	if (vp->tx_queue)
+		spin_lock(&vp->tx_queue->head_lock);
  	vp->estats.tx_timeout_count = 0;
  	vp->estats.tx_restart_queue = 0;
  	vp->estats.tx_kicks = 0;
@@ -131,7 +124,8 @@ static void vector_reset_stats(struct vector_private *vp)
  	vp->estats.tx_flow_control_xoff = 0;
  	vp->estats.tx_queue_max = 0;
  	vp->estats.tx_queue_running_average = 0;
-	spin_unlock(&vp->tx_queue->head_lock);
+	if (vp->tx_queue)
+		spin_unlock(&vp->tx_queue->head_lock);
  }
 
  static int get_mtu(struct arglist *def)
@@ -1163,7 +1157,8 @@ static int vector_poll(struct napi_struct *napi, int budget)
 
  	if ((vp->options & VECTOR_TX) != 0)
  		tx_enqueued = (vector_send(vp->tx_queue) > 0);
-	spin_lock(&vp->rx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_lock(&vp->rx_queue->head_lock);
  	if ((vp->options & VECTOR_RX) > 0)
  		err = vector_mmsg_rx(vp, budget);
  	else {
@@ -1171,7 +1166,8 @@ static int vector_poll(struct napi_struct *napi, int budget)
  		if (err > 0)
  			err = 1;
  	}
-	spin_unlock(&vp->rx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_unlock(&vp->rx_queue->head_lock);
  	if (err > 0)
  		work_done += err;
 
@@ -1421,10 +1417,10 @@ static void vector_get_ringparam(struct net_device *netdev,
  {
  	struct vector_private *vp = netdev_priv(netdev);
 
-	ring->rx_max_pending = vp->rx_queue->max_depth;
-	ring->tx_max_pending = vp->tx_queue->max_depth;
-	ring->rx_pending = vp->rx_queue->max_depth;
-	ring->tx_pending = vp->tx_queue->max_depth;
+	ring->rx_max_pending = vp->rx_queue ? vp->rx_queue->max_depth : 0;
+	ring->tx_max_pending = vp->tx_queue ? vp->tx_queue->max_depth : 0;
+	ring->rx_pending = ring->rx_max_pending;
+	ring->tx_pending = ring->tx_max_pending;
  }
 
  static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
@@ -1466,11 +1462,15 @@ static void vector_get_ethtool_stats(struct net_device *dev,
  	 * to date.
  	 */
 
-	spin_lock(&vp->tx_queue->head_lock);
-	spin_lock(&vp->rx_queue->head_lock);
+	if (vp->tx_queue)
+		spin_lock(&vp->tx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_lock(&vp->rx_queue->head_lock);

arch/um/drivers/vector_kern.c static void vector_get_ethtool_stats(struct net_device *dev, · line 1476

Quoted 10 lines
arch/um/drivers/vector_kern.c +4 -2
  	memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
-	spin_unlock(&vp->rx_queue->head_lock);
-	spin_unlock(&vp->tx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_unlock(&vp->rx_queue->head_lock);
+	if (vp->tx_queue)
+		spin_unlock(&vp->tx_queue->head_lock);
  }
 
  static int vector_get_coalesce(struct net_device *netdev,
Signature · 2 lines
Anton R. Ivanov
Cambridgegreys Limited. Registered in England. Company Number 10273661
maintainer
#2 lore raw
In reply to Michael Bommarito

Sorry, I didn't pay much attention to this before...

Quoted 9 lines · Michael Bommarito

TAP transport sets neither VECTOR_RX nor VECTOR_TX, so vector_net_open() never allocates rx_queue or tx_queue. HYBRID sets VECTOR_RX but not VECTOR_TX, so tx_queue is NULL there too.

vector_reset_stats(), vector_poll(), vector_get_ethtool_stats(), and vector_get_ringparam() unconditionally deref these queue pointers, causing a NULL pointer crash on SMP or with any lock debugging option.

Guard all queue pointer accesses with NULL checks.

I see how that fixes the crash, but maybe you could write a few words on why it's still correct?

arch/um/drivers/vector_kern.c static void vector_get_ethtool_stats(struct net_device *dev, · line 1469

Quoted 7 lines
arch/um/drivers/vector_kern.c +4 -2
-	spin_lock(&vp->tx_queue->head_lock);
-	spin_lock(&vp->rx_queue->head_lock);
+	if (vp->tx_queue)
+		spin_lock(&vp->tx_queue->head_lock);
+	if (vp->rx_queue)
+		spin_lock(&vp->rx_queue->head_lock);
 	memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));

I could imagine for example this memcpy() observing a torn write or something like that and getting strange results out?

Or is that just not a thing because UML is (still) mostly non-SMP?

Also I think there are related issues that wouldn't show up for a broken configuration, such as if create_queue() fails to allocate memory and we get an inconsistency between tx_queue / rx_queue pointers and VECTOR_TX / VECTOR_RX flags? Though I'll admit that seems highly unlikely.

johannes