On Thu, 2013-08-29 at 20:16 +0200, Sebastian Hesselbarth wrote:

On 08/29/13 15:35, Arnd Bergmann wrote:

quoted

On Tuesday 27 August 2013, Sebastian Hesselbarth wrote:

quoted

@@ -422,16 +419,8 @@ void __init v2m_dt_init_early(void)
                         pr_warning("vexpress: DT HBI (%x) is not matching "
                                         "hardware (%x)!\n", dt_hbi, hbi);
         }
-}
-
-static void __init v2m_dt_timer_init(void)
-{
-       of_clk_init(NULL);

-       clocksource_of_init();
-
-       versatile_sched_clock_init(vexpress_get_24mhz_clock_base(),
-                               24000000);
+       versatile_sched_clock_init(vexpress_get_24mhz_clock_base(), 24000000);
  }

You are moving versatile_sched_clock_init() ahead of clocksource_of_init(), which I suspect
won't work. Have you checked this?

"Checked" as in "Tested", no I haven't.

But non-DT v2m has it in v2m_init_early also, while v2m_sp804_init()
is called in v2m_timer_init().

That matches the above approach taken for DT v2m where
versatile_sched_clock_init() is now called from v2m_dt_init_early()
and clocksource_of_init() called from arch-wide .timer_init.

get_maintainer.pl did not spit out any additional maintainer except
Russell of course. You know someone who can test the above?

After adding of_clk_init(NULL) to time_init() things boot OK for me with
this patch. However, do we know that sched_clock is never going to get
read before time_init() has actually started the clock it reads? Are we
making things more fragile?

(My testing was on a 3.11-rc6 based Linaro kernel on TC2 because that's
what I had to hand. So it's not a vanilla mainline kernel and I have
other vexpress changes in my tree.)

-- 
Tixy