[PATCH] shmobile: fix map_io init usage

[linux@arm.linux.org.uk (Russell King - ARM Linux)]

On Mon, May 16, 2011 at 01:28:34PM +0900, Magnus Damm wrote:
> On Fri, May 13, 2011 at 10:56 AM, Jean-Christophe PLAGNIOL-VILLARD
> <plagnioj@jcrosoft.com> wrote:
> > On 18:28 Sun 01 May ? ? , Jean-Christophe PLAGNIOL-VILLARD wrote:
> >> switch early init to init_early and introduce soc map_io
> >>
> >> with this Patch we will not do any more early device setup during the map io
> >>
> >> Signed-off-by: Jean-Christophe PLAGNIOL-VILLARD <plagnioj@jcrosoft.com>
> >> Cc: Russell King <rmk+kernel@arm.linux.org.uk>
> >> Cc: Paul Mundt <lethal@linux-sh.org>
> >> Cc: Magnus Damm <magnus.damm@gmail.com>
> > ping
> 
> Why? The device init order is changed by this patch to setup the
> serial console later. I fail to see the benefit. Basically, the serial
> console setup is delayed to happen in ->init_early() instead of
> ->map_io(). In practice this means that we restrict ourselves to not
> be able to output debug messages on the serial console from a bunch of
> calls in setup_arch().

Because that's how its supposed to work.  map_io is for setting up the
device mappings.  Those mappings need flushing out to RAM on some CPUs
to be visible, and that's taken care of _after_ map_io has returned.

If the page table entries haven't found their way to RAM, accessing
devices in map_io will cause a fault, and produce a silently non-booting
kernel.  With a write-alloc writeback cache policy, this becomes more
likely than the default policy of a read-alloc writeback cache.

Platforms should not do their own cache and TLB maintainence in their
map_io callback to make it work either - should we need to change
things that means going through all platforms.