[PATCH] netif_rx from non interrupt context

[PATCH] netif_rx from non interrupt context

[Maksim Krasnyanskiy]

Hi Folks,

Generic function for the net drivers that call netif_rx from non interrupt context.
And TUN/TAP driver patch.

--- linux/include/linux/netdevice.h.old Mon Jul 30 11:37:27 2001
+++ linux/include/linux/netdevice.h     Mon Jul 30 11:48:32 2001
@@ -563,6 +563,19 @@
  
  extern int             netdev_nit;
  
+
+/* 
+ * netif_rx_ni -       post buffer to the network code from _non interrupt_ context.
+ *                     see net/core/dev.c for netif_rx description.
+ */
+static inline int netif_rx_ni(struct sk_buff *skb)
+{
+       int err = netif_rx(skb);
+       if (softirq_pending(smp_processor_id()))
+               do_softirq();
+       return err;
+}
+

--- linux/drivers/net/tun.c.old Mon Jun 11 19:15:27 2001
+++ linux/drivers/net/tun.c     Mon Jul 30 11:49:01 2001
@@ -218,7 +218,7 @@
         if (tun->flags & TUN_NOCHECKSUM)
                 skb->ip_summed = CHECKSUM_UNNECESSARY;
   
-       netif_rx(skb);
+       netif_rx_ni(skb);
     
         tun->stats.rx_packets++;
         tun->stats.rx_bytes += len;


Thanks
Max

Maksim Krasnyanskiy	
Senior Kernel Engineer
Qualcomm Incorporated

maxk@qualcomm.com
http://bluez.sf.net
http://vtun.sf.net

Re: [PATCH] netif_rx from non interrupt context

[David S. Miller]

Maksim Krasnyanskiy writes:
 > Generic function for the net drivers that call netif_rx from non interrupt context.
 > And TUN/TAP driver patch.

No, let us do it explicitly in the drivers, not create a new API for
this.

Maybe we should add "Send to socket with BH disabled" or a "insert to
generic linked list with spinlock held" interfaces too? :-)

Later,
David S. Miller
davem@redhat.com

Re: [PATCH] netif_rx from non interrupt context

[Maksim Krasnyanskiy]

>  > Generic function for the net drivers that call netif_rx from non interrupt context.
>  > And TUN/TAP driver patch.
>
>No, let us do it explicitly in the drivers, not create a new API for this.
Well, the thing is that we can probably optimize that function (like you guys did for a local_bh_enable) because it's
a critical path. Also it makes sense (to me) to hide softirq implementation details from the net drivers.
If softirqs are changed again we can fix one place instead of auditing net drivers.

>Maybe we should add "Send to socket with BH disabled" or a "insert to generic linked list with spinlock held" interfaces too ? :-)
:-) 

Max

Maksim Krasnyanskiy	
Senior Kernel Engineer
Qualcomm Incorporated

maxk@qualcomm.com
http://bluez.sf.net
http://vtun.sf.net

Re: [PATCH] netif_rx from non interrupt context

[kuznet]

Hello!

> a critical path. Also it makes sense (to me) to hide softirq implementation details from the net drivers.

He tells right thing... The fact that netif_rx() is invalid in this context
and, especially, way to make it valid is not evident. It was not evident
for me three days ago, at least. :-)

Seems, it is better to hide this yet.

Alexey