Re: [PATCH] net: fec: fix regression on i.MX28 introduced by rx_copybreak support

[Russell King - ARM Linux <linux@arm.linux.org.uk>]

On Wed, Oct 08, 2014 at 08:54:58AM +0000, David Laight wrote:
> Hmmm... in that case you may not want the compiler to convert the bit value
> to a 'bool' at all.
> 
> Passing 'id_entry->driver_data' through (that doesn't look like a field name for
> 'quirk flags) would generate better code.
> 
> Even better would be to reference the flag directly from 'ndev'.
> A pointer indirection for the test if probably faster then passing
> another argument.

A far better idea would be to copy the quirks into the fec_net_private
structure, storing them as a 'unsigned int' value, and test them from
there.  This is /much/ more efficient than jumping through the hoops to
retrieve id_entry, and then testing the 64-bit driver_data value.

I've had such a patch since about the beginning of the year (and patches
which add stuff like byte queue limits, which are really needed now that
we have a /huge/ transmit ring), but I can't keep up with rebasing the
patch set, and properly tested and performance impacts evaluated due to
the rate of FEC changes.  I never finished rebasing the set after Andy's
TSO work...

-- 
FTTC broadband for 0.8mile line: currently at 9.5Mbps down 400kbps up
according to speedtest.net.