part of commit 0dcffac1 upstream
(the upstream multislice patch contains this fix within a large
 rework of the code since there is one rx_done ring per slice. The
 old allocating is replaced by a call to myri10ge_probe_slices())


Allocating the rx_done ring requires mgp->max_intr_slots to
be valid, which requires that we already loaded the firmware
we are going to use with this ring.
So move the allocating after myri10ge_load_firmware()
(but keep it before myri10ge_reset() which already needs the
 rx_done ring).

If fixedsa regression where loading the driver would not
appear to do anything. Regression introduced in 2.6.26-rc3 by
commit 014377a1

Reported and patch tested by Lukas Hejtmanek at
http://lkml.org/lkml/2008/7/22/305
Reproduced and patch tested also by me.
Signed-off-by: Brice Goglin <brice@myri.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>