Linux Virtualization - Jul 2019 - [PATCH v3 1/3] vsock/virtio: use RCU to avoid use-after-free on the_virtio_vsock

During the review of "[PATCH] vsock/virtio: Initialize core virtio vsock
before registering the driver", Stefan pointed out some possible issues
in the .probe() and .remove() callbacks of the virtio-vsock driver.

This series tries to solve these issues:
- Patch 1 adds RCU critical sections to avoid use-after-free of
  'the_virtio_vsock' pointer.
- Patch 2 stops workers before to call vdev->config->reset(vdev) to
  be sure that no one is accessing the device.
- Patch 3 moves the works flush at the end of the .remove() to avoid
  use-after-free of 'vsock' object.

v3:
- Patch 1: use rcu_dereference_protected() to get the_virtio_vosck value in
           the virtio_vsock_probe() [Jason]

v2: https://patchwork.kernel.org/cover/11022343/

v1: https://patchwork.kernel.org/cover/10964733/

Before this series the guest crashes in a few second. After this series the
test runs (~12h) without issues.
Tested on an SMP guest (-smp 4 -monitor tcp:127.0.0.1:1234,server,nowait)
with these scripts to stress the .probe()/.remove() path:

- guest
  while true; do
      cat /dev/urandom | nc-vsock -l 4321 > /dev/null &
      cat /dev/urandom | nc-vsock -l 5321 > /dev/null &
      cat /dev/urandom | nc-vsock -l 6321 > /dev/null &
      cat /dev/urandom | nc-vsock -l 7321 > /dev/null &
      wait
  done

- host
  while true; do
      cat /dev/urandom | nc-vsock 3 4321 > /dev/null &
      cat /dev/urandom | nc-vsock 3 5321 > /dev/null &
      cat /dev/urandom | nc-vsock 3 6321 > /dev/null &
      cat /dev/urandom | nc-vsock 3 7321 > /dev/null &
      sleep 2
      echo "device_del v1" | nc 127.0.0.1 1234
      sleep 1
      echo "device_add vhost-vsock-pci,id=v1,guest-cid=3" | nc
127.0.0.1 1234
      sleep 1
  done

Stefano Garzarella (3):
  vsock/virtio: use RCU to avoid use-after-free on the_virtio_vsock
  vsock/virtio: stop workers during the .remove()
  vsock/virtio: fix flush of works during the .remove()

 net/vmw_vsock/virtio_transport.c | 134 ++++++++++++++++++++++++-------
 1 file changed, 104 insertions(+), 30 deletions(-)

-- 
2.20.1

Some callbacks used by the upper layers can run while we are in the
.remove(). A potential use-after-free can happen, because we free
the_virtio_vsock without knowing if the callbacks are over or not.

To solve this issue we move the assignment of the_virtio_vsock at the
end of .probe(), when we finished all the initialization, and at the
beginning of .remove(), before to release resources.
For the same reason, we do the same also for the vdev->priv.

We use RCU to be sure that all callbacks that use the_virtio_vsock
ended before freeing it. This is not required for callbacks that
use vdev->priv, because after the vdev->config->del_vqs() we are sure
that they are ended and will no longer be invoked.

We also take the mutex during the .remove() to avoid that .probe() can
run while we are resetting the device.

Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 net/vmw_vsock/virtio_transport.c | 70 +++++++++++++++++++++-----------
 1 file changed, 46 insertions(+), 24 deletions(-)

diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
index 9c287e3e393c..3eaec60aa64f 100644
--- a/net/vmw_vsock/virtio_transport.c
+++ b/net/vmw_vsock/virtio_transport.c
@@ -65,19 +65,22 @@ struct virtio_vsock {
 	u32 guest_cid;
 };
 
-static struct virtio_vsock *virtio_vsock_get(void)
-{
-	return the_virtio_vsock;
-}
-
 static u32 virtio_transport_get_local_cid(void)
 {
-	struct virtio_vsock *vsock = virtio_vsock_get();
+	struct virtio_vsock *vsock;
+	u32 ret;
 
-	if (!vsock)
-		return VMADDR_CID_ANY;
+	rcu_read_lock();
+	vsock = rcu_dereference(the_virtio_vsock);
+	if (!vsock) {
+		ret = VMADDR_CID_ANY;
+		goto out_rcu;
+	}
 
-	return vsock->guest_cid;
+	ret = vsock->guest_cid;
+out_rcu:
+	rcu_read_unlock();
+	return ret;
 }
 
 static void virtio_transport_loopback_work(struct work_struct *work)
@@ -197,14 +200,18 @@ virtio_transport_send_pkt(struct virtio_vsock_pkt *pkt)
 	struct virtio_vsock *vsock;
 	int len = pkt->len;
 
-	vsock = virtio_vsock_get();
+	rcu_read_lock();
+	vsock = rcu_dereference(the_virtio_vsock);
 	if (!vsock) {
 		virtio_transport_free_pkt(pkt);
-		return -ENODEV;
+		len = -ENODEV;
+		goto out_rcu;
 	}
 
-	if (le64_to_cpu(pkt->hdr.dst_cid) == vsock->guest_cid)
-		return virtio_transport_send_pkt_loopback(vsock, pkt);
+	if (le64_to_cpu(pkt->hdr.dst_cid) == vsock->guest_cid) {
+		len = virtio_transport_send_pkt_loopback(vsock, pkt);
+		goto out_rcu;
+	}
 
 	if (pkt->reply)
 		atomic_inc(&vsock->queued_replies);
@@ -214,6 +221,9 @@ virtio_transport_send_pkt(struct virtio_vsock_pkt *pkt)
 	spin_unlock_bh(&vsock->send_pkt_list_lock);
 
 	queue_work(virtio_vsock_workqueue, &vsock->send_pkt_work);
+
+out_rcu:
+	rcu_read_unlock();
 	return len;
 }
 
@@ -222,12 +232,14 @@ virtio_transport_cancel_pkt(struct vsock_sock *vsk)
 {
 	struct virtio_vsock *vsock;
 	struct virtio_vsock_pkt *pkt, *n;
-	int cnt = 0;
+	int cnt = 0, ret;
 	LIST_HEAD(freeme);
 
-	vsock = virtio_vsock_get();
+	rcu_read_lock();
+	vsock = rcu_dereference(the_virtio_vsock);
 	if (!vsock) {
-		return -ENODEV;
+		ret = -ENODEV;
+		goto out_rcu;
 	}
 
 	spin_lock_bh(&vsock->send_pkt_list_lock);
@@ -255,7 +267,11 @@ virtio_transport_cancel_pkt(struct vsock_sock *vsk)
 			queue_work(virtio_vsock_workqueue, &vsock->rx_work);
 	}
 
-	return 0;
+	ret = 0;
+
+out_rcu:
+	rcu_read_unlock();
+	return ret;
 }
 
 static void virtio_vsock_rx_fill(struct virtio_vsock *vsock)
@@ -565,7 +581,8 @@ static int virtio_vsock_probe(struct virtio_device *vdev)
 		return ret;
 
 	/* Only one virtio-vsock device per guest is supported */
-	if (the_virtio_vsock) {
+	if (rcu_dereference_protected(the_virtio_vsock,
+				lockdep_is_held(&the_virtio_vsock_mutex))) {
 		ret = -EBUSY;
 		goto out;
 	}
@@ -590,8 +607,6 @@ static int virtio_vsock_probe(struct virtio_device *vdev)
 	vsock->rx_buf_max_nr = 0;
 	atomic_set(&vsock->queued_replies, 0);
 
-	vdev->priv = vsock;
-	the_virtio_vsock = vsock;
 	mutex_init(&vsock->tx_lock);
 	mutex_init(&vsock->rx_lock);
 	mutex_init(&vsock->event_lock);
@@ -613,6 +628,9 @@ static int virtio_vsock_probe(struct virtio_device *vdev)
 	virtio_vsock_event_fill(vsock);
 	mutex_unlock(&vsock->event_lock);
 
+	vdev->priv = vsock;
+	rcu_assign_pointer(the_virtio_vsock, vsock);
+
 	mutex_unlock(&the_virtio_vsock_mutex);
 	return 0;
 
@@ -627,6 +645,12 @@ static void virtio_vsock_remove(struct virtio_device *vdev)
 	struct virtio_vsock *vsock = vdev->priv;
 	struct virtio_vsock_pkt *pkt;
 
+	mutex_lock(&the_virtio_vsock_mutex);
+
+	vdev->priv = NULL;
+	rcu_assign_pointer(the_virtio_vsock, NULL);
+	synchronize_rcu();
+
 	flush_work(&vsock->loopback_work);
 	flush_work(&vsock->rx_work);
 	flush_work(&vsock->tx_work);
@@ -666,12 +690,10 @@ static void virtio_vsock_remove(struct virtio_device
*vdev)
 	}
 	spin_unlock_bh(&vsock->loopback_list_lock);
 
-	mutex_lock(&the_virtio_vsock_mutex);
-	the_virtio_vsock = NULL;
-	mutex_unlock(&the_virtio_vsock_mutex);
-
 	vdev->config->del_vqs(vdev);
 
+	mutex_unlock(&the_virtio_vsock_mutex);
+
 	kfree(vsock);
 }
 
-- 
2.20.1

Before to call vdev->config->reset(vdev) we need to be sure that
no one is accessing the device, for this reason, we add new variables
in the struct virtio_vsock to stop the workers during the .remove().

This patch also add few comments before vdev->config->reset(vdev)
and vdev->config->del_vqs(vdev).

Suggested-by: Stefan Hajnoczi <stefanha at redhat.com>
Suggested-by: Michael S. Tsirkin <mst at redhat.com>
Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 net/vmw_vsock/virtio_transport.c | 51 +++++++++++++++++++++++++++++++-
 1 file changed, 50 insertions(+), 1 deletion(-)

diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
index 3eaec60aa64f..4dbdce7746bd 100644
--- a/net/vmw_vsock/virtio_transport.c
+++ b/net/vmw_vsock/virtio_transport.c
@@ -38,6 +38,7 @@ struct virtio_vsock {
 	 * must be accessed with tx_lock held.
 	 */
 	struct mutex tx_lock;
+	bool tx_run;
 
 	struct work_struct send_pkt_work;
 	spinlock_t send_pkt_list_lock;
@@ -53,6 +54,7 @@ struct virtio_vsock {
 	 * must be accessed with rx_lock held.
 	 */
 	struct mutex rx_lock;
+	bool rx_run;
 	int rx_buf_nr;
 	int rx_buf_max_nr;
 
@@ -60,6 +62,7 @@ struct virtio_vsock {
 	 * vqs[VSOCK_VQ_EVENT] must be accessed with event_lock held.
 	 */
 	struct mutex event_lock;
+	bool event_run;
 	struct virtio_vsock_event event_list[8];
 
 	u32 guest_cid;
@@ -94,6 +97,10 @@ static void virtio_transport_loopback_work(struct work_struct
*work)
 	spin_unlock_bh(&vsock->loopback_list_lock);
 
 	mutex_lock(&vsock->rx_lock);
+
+	if (!vsock->rx_run)
+		goto out;
+
 	while (!list_empty(&pkts)) {
 		struct virtio_vsock_pkt *pkt;
 
@@ -102,6 +109,7 @@ static void virtio_transport_loopback_work(struct
work_struct *work)
 
 		virtio_transport_recv_pkt(pkt);
 	}
+out:
 	mutex_unlock(&vsock->rx_lock);
 }
 
@@ -130,6 +138,9 @@ virtio_transport_send_pkt_work(struct work_struct *work)
 
 	mutex_lock(&vsock->tx_lock);
 
+	if (!vsock->tx_run)
+		goto out;
+
 	vq = vsock->vqs[VSOCK_VQ_TX];
 
 	for (;;) {
@@ -188,6 +199,7 @@ virtio_transport_send_pkt_work(struct work_struct *work)
 	if (added)
 		virtqueue_kick(vq);
 
+out:
 	mutex_unlock(&vsock->tx_lock);
 
 	if (restart_rx)
@@ -323,6 +335,10 @@ static void virtio_transport_tx_work(struct work_struct
*work)
 
 	vq = vsock->vqs[VSOCK_VQ_TX];
 	mutex_lock(&vsock->tx_lock);
+
+	if (!vsock->tx_run)
+		goto out;
+
 	do {
 		struct virtio_vsock_pkt *pkt;
 		unsigned int len;
@@ -333,6 +349,8 @@ static void virtio_transport_tx_work(struct work_struct
*work)
 			added = true;
 		}
 	} while (!virtqueue_enable_cb(vq));
+
+out:
 	mutex_unlock(&vsock->tx_lock);
 
 	if (added)
@@ -361,6 +379,9 @@ static void virtio_transport_rx_work(struct work_struct
*work)
 
 	mutex_lock(&vsock->rx_lock);
 
+	if (!vsock->rx_run)
+		goto out;
+
 	do {
 		virtqueue_disable_cb(vq);
 		for (;;) {
@@ -470,6 +491,9 @@ static void virtio_transport_event_work(struct work_struct
*work)
 
 	mutex_lock(&vsock->event_lock);
 
+	if (!vsock->event_run)
+		goto out;
+
 	do {
 		struct virtio_vsock_event *event;
 		unsigned int len;
@@ -484,7 +508,7 @@ static void virtio_transport_event_work(struct work_struct
*work)
 	} while (!virtqueue_enable_cb(vq));
 
 	virtqueue_kick(vsock->vqs[VSOCK_VQ_EVENT]);
-
+out:
 	mutex_unlock(&vsock->event_lock);
 }
 
@@ -620,12 +644,18 @@ static int virtio_vsock_probe(struct virtio_device *vdev)
 	INIT_WORK(&vsock->send_pkt_work, virtio_transport_send_pkt_work);
 	INIT_WORK(&vsock->loopback_work, virtio_transport_loopback_work);
 
+	mutex_lock(&vsock->tx_lock);
+	vsock->tx_run = true;
+	mutex_unlock(&vsock->tx_lock);
+
 	mutex_lock(&vsock->rx_lock);
 	virtio_vsock_rx_fill(vsock);
+	vsock->rx_run = true;
 	mutex_unlock(&vsock->rx_lock);
 
 	mutex_lock(&vsock->event_lock);
 	virtio_vsock_event_fill(vsock);
+	vsock->event_run = true;
 	mutex_unlock(&vsock->event_lock);
 
 	vdev->priv = vsock;
@@ -660,6 +690,24 @@ static void virtio_vsock_remove(struct virtio_device *vdev)
 	/* Reset all connected sockets when the device disappear */
 	vsock_for_each_connected_socket(virtio_vsock_reset_sock);
 
+	/* Stop all work handlers to make sure no one is accessing the device,
+	 * so we can safely call vdev->config->reset().
+	 */
+	mutex_lock(&vsock->rx_lock);
+	vsock->rx_run = false;
+	mutex_unlock(&vsock->rx_lock);
+
+	mutex_lock(&vsock->tx_lock);
+	vsock->tx_run = false;
+	mutex_unlock(&vsock->tx_lock);
+
+	mutex_lock(&vsock->event_lock);
+	vsock->event_run = false;
+	mutex_unlock(&vsock->event_lock);
+
+	/* Flush all device writes and interrupts, device will not use any
+	 * more buffers.
+	 */
 	vdev->config->reset(vdev);
 
 	mutex_lock(&vsock->rx_lock);
@@ -690,6 +738,7 @@ static void virtio_vsock_remove(struct virtio_device *vdev)
 	}
 	spin_unlock_bh(&vsock->loopback_list_lock);
 
+	/* Delete virtqueues and flush outstanding callbacks if any */
 	vdev->config->del_vqs(vdev);
 
 	mutex_unlock(&the_virtio_vsock_mutex);
-- 
2.20.1

This patch moves the flush of works after vdev->config->del_vqs(vdev),
because we need to be sure that no workers run before to free the
'vsock' object.

Since we stopped the workers using the [tx|rx|event]_run flags,
we are sure no one is accessing the device while we are calling
vdev->config->reset(vdev), so we can safely move the workers' flush.

Before the vdev->config->del_vqs(vdev), workers can be scheduled
by VQ callbacks, so we must flush them after del_vqs(), to avoid
use-after-free of 'vsock' object.

Suggested-by: Michael S. Tsirkin <mst at redhat.com>
Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 net/vmw_vsock/virtio_transport.c | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
index 4dbdce7746bd..0815d1357861 100644
--- a/net/vmw_vsock/virtio_transport.c
+++ b/net/vmw_vsock/virtio_transport.c
@@ -681,12 +681,6 @@ static void virtio_vsock_remove(struct virtio_device *vdev)
 	rcu_assign_pointer(the_virtio_vsock, NULL);
 	synchronize_rcu();
 
-	flush_work(&vsock->loopback_work);
-	flush_work(&vsock->rx_work);
-	flush_work(&vsock->tx_work);
-	flush_work(&vsock->event_work);
-	flush_work(&vsock->send_pkt_work);
-
 	/* Reset all connected sockets when the device disappear */
 	vsock_for_each_connected_socket(virtio_vsock_reset_sock);
 
@@ -741,6 +735,15 @@ static void virtio_vsock_remove(struct virtio_device *vdev)
 	/* Delete virtqueues and flush outstanding callbacks if any */
 	vdev->config->del_vqs(vdev);
 
+	/* Other works can be queued before 'config->del_vqs()', so we
flush
+	 * all works before to free the vsock object to avoid use after free.
+	 */
+	flush_work(&vsock->loopback_work);
+	flush_work(&vsock->rx_work);
+	flush_work(&vsock->tx_work);
+	flush_work(&vsock->event_work);
+	flush_work(&vsock->send_pkt_work);
+
 	mutex_unlock(&the_virtio_vsock_mutex);
 
 	kfree(vsock);
-- 
2.20.1

From: Stefano Garzarella <sgarzare at redhat.com>
Date: Fri,  5 Jul 2019 13:04:51 +0200
> During the review of "[PATCH] vsock/virtio: Initialize core virtio
vsock
> before registering the driver", Stefan pointed out some possible
issues
> in the .probe() and .remove() callbacks of the virtio-vsock driver.
 ...

Series applied to net-next, thanks.