Re: [PATCH] virtio_ring: Fix data races in split virtqueue used ring accesses

[Michael S. Tsirkin]

On Thu, Mar 05, 2026 at 05:29:27PM +0800, Chaohai Chen wrote:
> KCSAN detected multiple data races when accessing the split virtqueue's
> used ring, which is shared memory concurrently accessed by both the CPU
> and the virtio device (hypervisor).
> 
> The races occur when reading the following fields without proper atomic
> operations:
> - vring.used->idx
> - vring.used->flags
> - vring.used->ring[].id
> - vring.used->ring[].len
> 
> These fields reside in DMA-shared memory and can be modified by the
> virtio device at any time. Without READ_ONCE(), the compiler may perform
> unsafe optimizations such as value caching or load tearing.

.... but does not.


> Example KCSAN report:
> 
> [  109.277250] ==================================================================
> [  109.283600] BUG: KCSAN: data-race in virtqueue_enable_cb_delayed_split+0x10f/0x170
> 
> [  109.295263] race at unknown origin, with read to 0xffff8b2a92ef2042 of 2 bytes by interrupt on cpu 1:
> [  109.306934]  virtqueue_enable_cb_delayed_split+0x10f/0x170
> [  109.312880]  virtqueue_enable_cb_delayed+0x3b/0x70
> [  109.318852]  start_xmit+0x315/0x860 [virtio_net]
> [  109.324532]  dev_hard_start_xmit+0x85/0x380
> [  109.329993]  sch_direct_xmit+0xd3/0x680
> [  109.335360]  __dev_xmit_skb+0x4ee/0xcc0
> [  109.340568]  __dev_queue_xmit+0x560/0xe00
> [  109.345701]  ip_finish_output2+0x49a/0x9b0
> [  109.350743]  __ip_finish_output+0x131/0x250
> [  109.355789]  ip_finish_output+0x28/0x180
> [  109.360712]  ip_output+0xa0/0x1c0
> [  109.365479]  __ip_queue_xmit+0x68d/0x9e0
> [  109.370156]  ip_queue_xmit+0x33/0x40
> [  109.374783]  __tcp_transmit_skb+0x1703/0x1970
> [  109.379467]  __tcp_send_ack.part.0+0x1bb/0x320
> ...
> [  109.499585]  do_idle+0x7a/0xe0
> [  109.502979]  cpu_startup_entry+0x25/0x30
> [  109.506481]  start_secondary+0x116/0x150
> [  109.509930]  common_startup_64+0x13e/0x141
> 
> [  109.516626] value changed: 0x0029 -> 0x002a
> 
> Fix these races by wrapping all reads from the used ring with READ_ONCE()
> to ensure:
> 1. The compiler always loads values from memory (no caching)
> 2. Loads are atomic (no load tearing)
> 3. The concurrent access intent is documented for KCSAN and developers
> 
> The changes affect the following functions:
> - virtqueue_kick_prepare_split(): used->flags and avail event
> - virtqueue_get_buf_ctx_split(): used->ring[].id and used->ring[].len
> - virtqueue_get_buf_ctx_split_in_order(): used->ring[].id and
> used->ring[].len
> - virtqueue_enable_cb_delayed_split(): used->idx
> 

These are no races, these are KCSAN false positives.
I am not against  documenting things using these macros
but i am against confusing commit log messages.



> Signed-off-by: Chaohai Chen <wdhh6@xxxxxxxxxx>
> ---
>  drivers/virtio/virtio_ring.c | 17 +++++++++--------
>  1 file changed, 9 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> index 335692d41617..a792a3f05837 100644
> --- a/drivers/virtio/virtio_ring.c
> +++ b/drivers/virtio/virtio_ring.c
> @@ -810,10 +810,10 @@ static bool virtqueue_kick_prepare_split(struct vring_virtqueue *vq)
>  
>  	if (vq->event) {
>  		needs_kick = vring_need_event(virtio16_to_cpu(vq->vq.vdev,
> -					vring_avail_event(&vq->split.vring)),
> +					READ_ONCE(vring_avail_event(&vq->split.vring))),
>  					      new, old);
>  	} else {
> -		needs_kick = !(vq->split.vring.used->flags &
> +		needs_kick = !(READ_ONCE(vq->split.vring.used->flags) &
>  					cpu_to_virtio16(vq->vq.vdev,
>  						VRING_USED_F_NO_NOTIFY));
>  	}
> @@ -940,9 +940,9 @@ static void *virtqueue_get_buf_ctx_split(struct vring_virtqueue *vq,
>  
>  	last_used = (vq->last_used_idx & (vq->split.vring.num - 1));
>  	i = virtio32_to_cpu(vq->vq.vdev,
> -			vq->split.vring.used->ring[last_used].id);
> +			READ_ONCE(vq->split.vring.used->ring[last_used].id));
>  	*len = virtio32_to_cpu(vq->vq.vdev,
> -			vq->split.vring.used->ring[last_used].len);
> +			READ_ONCE(vq->split.vring.used->ring[last_used].len));
>  
>  	if (unlikely(i >= vq->split.vring.num)) {
>  		BAD_RING(vq, "id %u out of range\n", i);
> @@ -1004,9 +1004,9 @@ static void *virtqueue_get_buf_ctx_split_in_order(struct vring_virtqueue *vq,
>  		virtio_rmb(vq->weak_barriers);
>  
>  		vq->batch_last.id = virtio32_to_cpu(vq->vq.vdev,
> -				    vq->split.vring.used->ring[last_used_idx].id);
> +				    READ_ONCE(vq->split.vring.used->ring[last_used_idx].id));
>  		vq->batch_last.len = virtio32_to_cpu(vq->vq.vdev,
> -				     vq->split.vring.used->ring[last_used_idx].len);
> +				     READ_ONCE(vq->split.vring.used->ring[last_used_idx].len));
>  	}
>  
>  	if (vq->batch_last.id == last_used) {
> @@ -1112,8 +1112,9 @@ static bool virtqueue_enable_cb_delayed_split(struct vring_virtqueue *vq)
>  			&vring_used_event(&vq->split.vring),
>  			cpu_to_virtio16(vq->vq.vdev, vq->last_used_idx + bufs));
>  
> -	if (unlikely((u16)(virtio16_to_cpu(vq->vq.vdev, vq->split.vring.used->idx)
> -					- vq->last_used_idx) > bufs)) {
> +	if (unlikely((u16)(virtio16_to_cpu(vq->vq.vdev,
> +				READ_ONCE(vq->split.vring.used->idx))
> +				- vq->last_used_idx) > bufs)) {
>  		END_USE(vq);
>  		return false;



I also want to know what this does to performance, or at least
code size.

>  	}
> -- 
> 2.43.7