On 04/24/2015 12:03 PM, David Laight wrote:
Sent: 22 April 2015 22:39
On 04/22/2015 11:42 PM, David Miller wrote:
Hmm, I've been digging in the net core, and was unable to see where TX
skb's get their NET_IP_ALIGN bytes reserved. Have I missed something?
Probably need to print out skb's fields...
NET_IP_ALIGN is for receive, not transmit.
But when I print 'skb->data' from the ndo_start_xmit() method (in the
'sh_eth' driver), all addresses end with 2, so it looks like NET_IP_ALIGN gets
added somewhere...
For a locally generated message:
The TCP userdata is likely to be 4 byte aligned.
The TCP and IP headers are multiples of 4 bytes.
The MAC header is 14 bytes.
So you end up with a buffer that starts on a 4n+2 boundary or an initial
short fragment that is 4n+2 bytes long.
If a message is being forwarded the alignment probably depends on where
it came from.
Thanks for the detailed reply, though it came a bit late. :-)
If you have ethernet hardware that requires tx or rx buffers to be on
The RX buffers can be adjusted with skb_resrerve(), it's only the TX
buffers that need to be copied...
4n boundaries you should send it back as 'not fit for purpose'.
I'm afraid we can't. :-)
However, my colleague has suggested a scheme minimizing the copying:
only up to 3 first bytes need to be copied to the driver's internal buffers,
the rest can be sent from an skb itself. That would require substantial
changes to the driver though...
David
WBR, Sergei
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
