Re: [syzbot] [kvm?] [net?] [virt?] general protection fault in virtio_transport_purge_skbs

[Arseniy Krasnov]

On 24.03.2023 12:06, Stefano Garzarella wrote:
> On Fri, Mar 24, 2023 at 9:55 AM Stefano Garzarella <sgarzare@xxxxxxxxxx> wrote:
>>
>> On Fri, Mar 24, 2023 at 9:31 AM Stefano Garzarella <sgarzare@xxxxxxxxxx> wrote:
>>>
>>> Hi Bobby,
>>> can you take a look at this report?
>>>
>>> It seems related to the changes we made to support skbuff.
>>
>> Could it be a problem of concurrent access to pkt_queue ?
>>
>> IIUC we should hold pkt_queue.lock when we call skb_queue_splice_init()
>> and remove pkt_list_lock. (or hold pkt_list_lock when calling
>> virtio_transport_purge_skbs, but pkt_list_lock seems useless now that
>> we use skbuff)
>>
> 
> In the previous patch was missing a hunk, new one attached:
> 
> #syz test https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git fff5a5e7f528
> 
> --- a/net/vmw_vsock/vsock_loopback.c
> +++ b/net/vmw_vsock/vsock_loopback.c
> @@ -15,7 +15,6 @@
>  struct vsock_loopback {
>         struct workqueue_struct *workqueue;
> 
> -       spinlock_t pkt_list_lock; /* protects pkt_list */
>         struct sk_buff_head pkt_queue;
>         struct work_struct pkt_work;
>  };
> @@ -32,9 +31,7 @@ static int vsock_loopback_send_pkt(struct sk_buff *skb)
>         struct vsock_loopback *vsock = &the_vsock_loopback;
>         int len = skb->len;
> 
> -       spin_lock_bh(&vsock->pkt_list_lock);
>         skb_queue_tail(&vsock->pkt_queue, skb);
Hello Stefano and Bobby,

Small remark, may be here we can use virtio_vsock_skb_queue_tail() instead of skb_queue_tail().
skb_queue_tail() disables irqs during spinlock access, while  virtio_vsock_skb_queue_tail()
uses spin_lock_bh(). vhost and virtio transports use virtio_vsock_skb_queue_tail().

Thanks, Arseniy
> -       spin_unlock_bh(&vsock->pkt_list_lock);
> 
>         queue_work(vsock->workqueue, &vsock->pkt_work);
> 
> @@ -113,9 +110,9 @@ static void vsock_loopback_work(struct work_struct *work)
> 
>         skb_queue_head_init(&pkts);
> 
> -       spin_lock_bh(&vsock->pkt_list_lock);
> +       spin_lock_bh(&vsock->pkt_queue.lock);
>         skb_queue_splice_init(&vsock->pkt_queue, &pkts);
> -       spin_unlock_bh(&vsock->pkt_list_lock);
> +       spin_unlock_bh(&vsock->pkt_queue.lock);
> 
>         while ((skb = __skb_dequeue(&pkts))) {
>                 virtio_transport_deliver_tap_pkt(skb);
> @@ -132,7 +129,6 @@ static int __init vsock_loopback_init(void)
>         if (!vsock->workqueue)
>                 return -ENOMEM;
> 
> -       spin_lock_init(&vsock->pkt_list_lock);
>         skb_queue_head_init(&vsock->pkt_queue);
>         INIT_WORK(&vsock->pkt_work, vsock_loopback_work);
> 
> @@ -156,9 +152,7 @@ static void __exit vsock_loopback_exit(void)
> 
>         flush_work(&vsock->pkt_work);
> 
> -       spin_lock_bh(&vsock->pkt_list_lock);
>         virtio_vsock_skb_queue_purge(&vsock->pkt_queue);
> -       spin_unlock_bh(&vsock->pkt_list_lock);
> 
>         destroy_workqueue(vsock->workqueue);
>  }
>