[PATCH net] vhost/net: complete zerocopy ubufs only once

[PATCH net] vhost/net: complete zerocopy ubufs only once

[Qing Ming]

vhost-net initializes one ubuf_info per outstanding zerocopy TX
descriptor and hands it to the backend socket.  The networking stack may
then clone a zerocopy skb before all skb references are released.  For
example, batman-adv fragmentation reaches skb_split(), which calls
skb_zerocopy_clone() and increments the same ubuf_info refcount.

vhost_zerocopy_complete() currently treats every ubuf callback as a
completed vhost descriptor.  It dereferences ubuf->ctx, writes the
descriptor completion state, and drops the vhost_net_ubuf_ref even when
the callback only releases a cloned skb reference.  A backend reset can
therefore wait for and free the vhost_net_ubuf_ref while another cloned
skb still carries the same ubuf_info.  A later completion then
dereferences the freed ubufs pointer.

KASAN reports the stale completion as:

  BUG: KASAN: slab-use-after-free in vhost_zerocopy_complete+0x1d7/0x1f0
  BUG: KASAN: slab-use-after-free in vhost_zerocopy_complete+0x101/0x1f0
  vhost_zerocopy_complete
  skb_copy_ubufs
  __dev_forward_skb2
  veth_xmit

The freed object was allocated from vhost_net_ioctl() while setting the
backend and freed through kfree_rcu()/kvfree_rcu_bulk after backend
removal, while delayed skb completion still reached
vhost_zerocopy_complete().

Honor the generic ubuf_info refcount before touching vhost state, and run
the vhost descriptor completion only for the final ubuf reference.  This
matches the msg_zerocopy_complete() ownership rule for cloned zerocopy
skbs.

Fixes: bab632d69ee4 ("vhost: vhost TX zero-copy support")
Signed-off-by: Qing Ming <a0yami@mailbox.org>
---
 drivers/vhost/net.c | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
index c6536cad9c4f..b9af63fb6306 100644
--- a/drivers/vhost/net.c
+++ b/drivers/vhost/net.c
@@ -390,13 +390,20 @@ static void vhost_zerocopy_signal_used(struct vhost_net *net,
 static void vhost_zerocopy_complete(struct sk_buff *skb,
 				    struct ubuf_info *ubuf_base, bool success)
 {
-	struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
-	struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
-	struct vhost_virtqueue *vq = ubufs->vq;
+	struct ubuf_info_msgzc *ubuf;
+	struct vhost_net_ubuf_ref *ubufs;
+	struct vhost_virtqueue *vq;
 	int cnt;
 
-	rcu_read_lock_bh();
+	/* Only the final cloned skb reference completes the vhost descriptor. */
+	if (!refcount_dec_and_test(&ubuf_base->refcnt))
+		return;
+
+	ubuf = uarg_to_msgzc(ubuf_base);
+	ubufs = ubuf->ctx;
+	vq = ubufs->vq;
 
+	rcu_read_lock_bh();
 	/* set len to mark this desc buffers done DMA */
 	vq->heads[ubuf->desc].len = success ?
 		VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
-- 
2.53.0

Re: [PATCH net] vhost/net: complete zerocopy ubufs only once

[Paolo Abeni]

On 6/1/26 12:43 PM, Qing Ming wrote:
> vhost-net initializes one ubuf_info per outstanding zerocopy TX
> descriptor and hands it to the backend socket.  The networking stack may
> then clone a zerocopy skb before all skb references are released.  For
> example, batman-adv fragmentation reaches skb_split(), which calls
> skb_zerocopy_clone() and increments the same ubuf_info refcount.
> 
> vhost_zerocopy_complete() currently treats every ubuf callback as a
> completed vhost descriptor.  It dereferences ubuf->ctx, writes the
> descriptor completion state, and drops the vhost_net_ubuf_ref even when
> the callback only releases a cloned skb reference.  A backend reset can
> therefore wait for and free the vhost_net_ubuf_ref while another cloned
> skb still carries the same ubuf_info.  A later completion then
> dereferences the freed ubufs pointer.
> 
> KASAN reports the stale completion as:
> 
>   BUG: KASAN: slab-use-after-free in vhost_zerocopy_complete+0x1d7/0x1f0
>   BUG: KASAN: slab-use-after-free in vhost_zerocopy_complete+0x101/0x1f0
>   vhost_zerocopy_complete
>   skb_copy_ubufs
>   __dev_forward_skb2
>   veth_xmit
> 
> The freed object was allocated from vhost_net_ioctl() while setting the
> backend and freed through kfree_rcu()/kvfree_rcu_bulk after backend
> removal, while delayed skb completion still reached
> vhost_zerocopy_complete().
> 
> Honor the generic ubuf_info refcount before touching vhost state, and run
> the vhost descriptor completion only for the final ubuf reference.  This
> matches the msg_zerocopy_complete() ownership rule for cloned zerocopy
> skbs.
> 
> Fixes: bab632d69ee4 ("vhost: vhost TX zero-copy support")
> Signed-off-by: Qing Ming <a0yami@mailbox.org>

The patch LGTM.

@Michael: to you want to take it via your tree?

/P

Re: [PATCH net] vhost/net: complete zerocopy ubufs only once

[Michael S. Tsirkin]

On Thu, Jun 04, 2026 at 12:55:49PM +0200, Paolo Abeni wrote:
> On 6/1/26 12:43 PM, Qing Ming wrote:
> > vhost-net initializes one ubuf_info per outstanding zerocopy TX
> > descriptor and hands it to the backend socket.  The networking stack may
> > then clone a zerocopy skb before all skb references are released.  For
> > example, batman-adv fragmentation reaches skb_split(), which calls
> > skb_zerocopy_clone() and increments the same ubuf_info refcount.
> > 
> > vhost_zerocopy_complete() currently treats every ubuf callback as a
> > completed vhost descriptor.  It dereferences ubuf->ctx, writes the
> > descriptor completion state, and drops the vhost_net_ubuf_ref even when
> > the callback only releases a cloned skb reference.  A backend reset can
> > therefore wait for and free the vhost_net_ubuf_ref while another cloned
> > skb still carries the same ubuf_info.  A later completion then
> > dereferences the freed ubufs pointer.
> > 
> > KASAN reports the stale completion as:
> > 
> >   BUG: KASAN: slab-use-after-free in vhost_zerocopy_complete+0x1d7/0x1f0
> >   BUG: KASAN: slab-use-after-free in vhost_zerocopy_complete+0x101/0x1f0
> >   vhost_zerocopy_complete
> >   skb_copy_ubufs
> >   __dev_forward_skb2
> >   veth_xmit
> > 
> > The freed object was allocated from vhost_net_ioctl() while setting the
> > backend and freed through kfree_rcu()/kvfree_rcu_bulk after backend
> > removal, while delayed skb completion still reached
> > vhost_zerocopy_complete().
> > 
> > Honor the generic ubuf_info refcount before touching vhost state, and run
> > the vhost descriptor completion only for the final ubuf reference.  This
> > matches the msg_zerocopy_complete() ownership rule for cloned zerocopy
> > skbs.
> > 
> > Fixes: bab632d69ee4 ("vhost: vhost TX zero-copy support")
> > Signed-off-by: Qing Ming <a0yami@mailbox.org>
> 
> The patch LGTM.
> 
> @Michael: to you want to take it via your tree?
> 
> /P


I wan't copied) Alright then.