Regarding net_device_ops

[Rami Rosen]

Hi, Pranay,

First, let's assume that you are talking about IPv4, though you did not
mention it explicitly. (The principles in IPv6 are quite similar, though)

A packet is sent out in the usual case with the ip_queue_xmit() method.
The  ip_queue_xmit() method calls the ip_route_output_ports() method
in order to perform a lookup in the IPv4 routing tables.

see: http://lxr.free-electrons.com/source/net/ipv4/ip_output.c#L352

The results of this IPv4 routing lookup determines on which network device
(net_device) the packet will be sent. You should look at the code of
ip_route_output_ports() method in net/ipv4/route.c in order to
understand the IPv4 routing subsystem and the IPv4 routing lookup.

Packets can, under certain circumstances, be sent by the ip_send_skb()
method, but this happens when the flow (which consists also of the
net_device to be used) is known before.

Best Regards,
Rami Rosen
http://ramirose.wix.com/ramirosen


On Sat, May 3, 2014 at 12:43 PM, Pranay Srivastava <pranjas at gmail.com> wrote:
> Hi
>
> Referring to Documentation/networking/netdevices.txt
>
> a)
>
> ndo->ndo_start_xmit if written "as is" will be thread safe? [Correct?]
> NETIF_F_LLTX is not set only then this is true.
>
> But dev_queue_xmit(skb) doesn't seem to take any lock, neither does
> dev_hard_start_xmit. Please let me know if i'm wrong and where is this
> lock taken before calling dev_queue_xmit/dev_hard_start_xmit.
>
>
> b)
>
> Why and when would I want to use rtnl_lock/unlock? Is it like
> registering multiple device simultaneously?
>
>
> c)
>
> ndo->ndo_start_xmit is supposed to start the transmission of the skb.
> However I've a doubt
>
> ---> Suppose that there's no room left for the incoming skb in the
> device buffer, so in that case it would be ideal to call
> netif_stop_queue. At this the the skb already with the ndo_start_xmit
> will not be retried again[Correct?]
>
> ----> Is ndo_start_xmit required to wait in this case? However if the
> above spin_lock in a) was taken then this would be a waste. So is it
> possible that this skb is "deffered" to whenever there's room and
> return NETDEV_TX_OK? But if we do this we are telling the application
> that it's packet was delivered so in case there was a timeout set by
> the application on a reply it would again send us the same packet. So
> instead of "deferring" is it good to "drop" ?
>
>
>
> --
>         ---P.K.S
>
> _______________________________________________
> Kernelnewbies mailing list
> Kernelnewbies at kernelnewbies.org
> http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies