On Mon, 2012-06-04 at 16:12 -0400, Chris Metcalf wrote:

This change adds support for the tilegx network driver based on the
GXIO IORPC support in the tilegx software stack, using the on-chip
mPIPE packet processing engine.

+
+/* Do "TSO" handling for egress.
+ *
+ * Normally drivers set NETIF_F_TSO only to support hardware TSO;
+ * otherwise the stack uses scatter-gather to implement GSO in software.
+ * On our testing, enabling GSO support (via NETIF_F_SG) drops network
+ * performance down to around 7.5 Gbps on the 10G interfaces, although
+ * also dropping cpu utilization way down, to under 8%.  But
+ * implementing "TSO" in the driver brings performance back up to line
+ * rate, while dropping cpu usage even further, to less than 4%.  In
+ * practice, profiling of GSO shows that skb_segment() is what causes
+ * the performance overheads; we benefit in the driver from using
+ * preallocated memory to duplicate the TCP/IP headers.
+ */

All this stuff cost about 300 lines of code in this driver, without IPv6
support.

I am pretty sure this performance problem should be solved in net/{core|
ipv4|ipv6} instead

What TCP performance do you get with TSO/GSO and SG off ?