Re: [PATCH] xen-netfront: handle NULL returned by xdp_convert_buff_to_frame()

[Alexey <sdl@xxxxxxxx>]

On 17.04.2025 11:51, Juergen Gross wrote:

> On
>       17.04.25 10:45, Alexey wrote:
>       
> >         
> >         On 17.04.2025 10:12, Jürgen Groß wrote:
> >         
> > > On 17.04.25 09:00, Alexey wrote:
> > >           
> > > >             
> > > >             On 17.04.2025 03:58, Jakub Kicinski wrote:
> > > >             
> > > > > On Mon, 14 Apr 2025 18:34:01 +0000
> > > > >               Alexey Nepomnyashih wrote:
> > > > >               
> > > > > >           get_page(pdata);
> > > > > >                 
> > > > > >               
> > > > >               Please notice this get_page() here.
> > > > >               
> > > > >               
> > > > > >           xdpf =
> > > > > >                 xdp_convert_buff_to_frame(xdp);
> > > > > >                 
> > > > > >                 +        if (unlikely(!xdpf)) {
> > > > > >                 
> > > > > >                 +           
> > > > > >                 trace_xdp_exception(queue->info->netdev, prog,
> > > > > >                 act);
> > > > > >                 
> > > > > >                 +            break;
> > > > > >                 
> > > > > >                 +        }
> > > > > >                 
> > > > > >               
> > > > >             
> > > >             Do you mean that it would be better to move the
> > > >             get_page(pdata) call lower,
> > > >             
> > > >             after checking for NULL in xdpf, so that the reference count
> > > >             is only increased
> > > >             
> > > >             after a successful conversion?
> > > >             
> > > >           
> > >           
> > >           I think the error handling here is generally broken (or at
> > >           least very
> > >           
> > >           questionable).
> > >           
> > >           
> > >           I suspect that in case of at least some errors the get_page()
> > >           is leaking
> > >           
> > >           even without this new patch.
> > >           
> > >           
> > >           In case I'm wrong a comment reasoning why there is no leak
> > >           should be
> > >           
> > >           added.
> > >           
> > >           
> > >           
> > >           Juergen
> > >           
> > >         
> >         
> >         I think pdata is freed in xdp_return_frame_rx_napi() ->
> >         __xdp_return()
> >         
> >       
>       
>       Agreed. But what if xennet_xdp_xmit() returns an error < 0?
>       
>       
>       In this case xdp_return_frame_rx_napi() won't be called.
>       
>       
>       
>       Juergen
>       
>

Agreed. There is no explicit freed pdata in the calling function
xennet_get_responses(). Without this, the page referenced by pdata
could be leaked.

I suggest:

 	case XDP_TX:
-		get_page(pdata);
 		xdpf = xdp_convert_buff_to_frame(xdp);
+		if (unlikely(!xdpf)) {
+			trace_xdp_exception(queue->info->netdev, prog, act);
+			break;
+		}
+		get_page(pdata);
 		err = xennet_xdp_xmit(queue->info->netdev, 1, &xdpf, 0);
 		if (unlikely(!err))
 			xdp_return_frame_rx_napi(xdpf);
-		else if (unlikely(err < 0))
+		else if (unlikely(err < 0)) {
 			trace_xdp_exception(queue->info->netdev, prog, act);
+			xdp_return_frame_rx_napi(xdpf);
+		}
 		break;
 	case XDP_REDIRECT:
 		get_page(pdata);
 		err = xdp_do_redirect(queue->info->netdev, xdp, prog);
 		*need_xdp_flush = true;
-		if (unlikely(err))
+		if (unlikely(err)) {
 			trace_xdp_exception(queue->info->netdev, prog, act);
+			__xdp_return(page_address(pdata), &xdp->mem, true, xdp);
+		}
 		break;