Re: [Qemu-devel] virtio_stw_phys_cached crash

[Fam Zheng]

On Wed, 08/22 19:45, Paolo Bonzini wrote:
> On 22/08/2018 16:57, Fam Zheng wrote:
> > On master (13b7b188501) and v3.0.0, booting the OpenBSD install63.iso with a
> > virtio-net (slirp backend) device gives this assertion failure:
> > 
> > (gdb) bt
> > #0  0x00007f6f25703feb in raise () at /lib64/libc.so.6
> > #1  0x00007f6f256ee5c1 in abort () at /lib64/libc.so.6
> > #2  0x00007f6f256ee491 in _nl_load_domain.cold.0 () at /lib64/libc.so.6
> > #3  0x00007f6f256fc752 in  () at /lib64/libc.so.6
> > #4  0x000055d966f5b5d6 in address_space_stw_le_cached 
> > (cache=0x7f6f040fd700, addr=516, val=1, attrs=..., result=0x0)
> >     at /stor/work/qemu/include/exec/memory_ldst_cached.inc.h:85
> 
> Does this help?

Nope. The crash happens the same (once you press "Auto Install" at the first
prompt).

Fam

> 
> diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c
> index d4e4d98b59..5982678c75 100644
> --- a/hw/virtio/virtio.c
> +++ b/hw/virtio/virtio.c
> @@ -2006,14 +2006,25 @@ static int virtio_set_features_nocheck(VirtIODevice 
> *vdev, uint64_t val)
>  
>  int virtio_set_features(VirtIODevice *vdev, uint64_t val)
>  {
> -   /*
> +    int ret;
> +    /*
>       * The driver must not attempt to set features after feature negotiation
>       * has finished.
>       */
>      if (vdev->status & VIRTIO_CONFIG_S_FEATURES_OK) {
>          return -EINVAL;
>      }
> -    return virtio_set_features_nocheck(vdev, val);
> +    ret = virtio_set_features_nocheck(vdev, val);
> +    if (!ret && (val & VIRTIO_RING_F_EVENT_IDX)) {
> +        /* VIRTIO_RING_F_EVENT_IDX changes the size of the caches.  */
> +        int i;
> +        for (i = 0; i < VIRTIO_QUEUE_MAX; i++) {
> +            if (vdev->vq[i].vring.num != 0) {
> +                virtio_init_region_cache(vdev, i);
> +            }
> +        }
> +    }
> +    return ret;
>  }
>  
>  int virtio_load(VirtIODevice *vdev, QEMUFile *f, int version_id)
> 
> (I haven't tried to reproduce, or checked the spec to see if this could
> be a guest bug too.  Of course assertion failures are wrong anyway, so
> we might as well work around it as above).
> 
> Paolo