[PATCH] arm64: kernel: compiling issue, need 'EXPORT_SYMBOL_GPL(read_current_timer)'

[marc.zyngier@arm.com (Marc Zyngier)]

On Tue, 21 May 2013 12:06:52 +0800, Chen Gang <gang.chen@asianux.com>
wrote:
> On 05/20/2013 05:56 PM, Will Deacon wrote:
>> On Mon, May 20, 2013 at 08:15:04AM +0100, Marc Zyngier wrote:
>>> On Mon, 20 May 2013 14:48:05 +0800, Chen Gang <gang.chen@asianux.com>
>>> wrote:
>>>> Need 'EXPORT_SYMBOL_GPL(read_current_timer)' if build with
>>>> allmodconfig.
>>>>
>>>> The related error:
>>>>   ERROR: "read_current_timer" [lib/rbtree_test.ko] undefined!
>>>>   ERROR: "read_current_timer" [lib/interval_tree_test.ko] undefined!
>>>>   ERROR: "read_current_timer" [fs/ext4/ext4.ko] undefined!
>>>>   ERROR: "read_current_timer" [crypto/tcrypt.ko] undefined!
>>>>
>>>>
>>>> Signed-off-by: Chen Gang <gang.chen@asianux.com>
>>>> ---
>>>>  arch/arm64/kernel/time.c |    1 +
>>>>  1 files changed, 1 insertions(+), 0 deletions(-)
>>>>
>>>> diff --git a/arch/arm64/kernel/time.c b/arch/arm64/kernel/time.c
>>>> index a551f88..7fcba80 100644
>>>> --- a/arch/arm64/kernel/time.c
>>>> +++ b/arch/arm64/kernel/time.c
>>>> @@ -73,6 +73,7 @@ int read_current_timer(unsigned long *timer_value)
>>>>  	*timer_value = arch_timer_read_counter();
>>>>  	return 0;
>>>>  }
>>>> +EXPORT_SYMBOL_GPL(read_current_timer);
>>>>  
>>>>  void __init time_init(void)
>>>>  {
>>>
>>> While this solves the problem, I'm not sure this is the best fix. The
>>> real
>>> issue is with get_cycles, which is a macro around read_current_timer.
>>>
>>> AArch32 exports it because of the number of timer implementations. On
>>> arm64, we should be able to just return CNTVCT_EL0.
>>>
>>> Catalin, Will, what do you think?
>> 
>> Should be ok once the arch timer driver has moved exclusively to
virtual
>> time. I'm also not sure we even need to implement read_current_timer()
--
>> it's only used for delay-loop calibration, which we don't need for the
>> arch timer.
>> 
> 
> For whether we need implement read_current_timer():
> 
>   many platforms have implemented it (openrisc, arm, sparc, hexagon,
>   avr32, x86).
>   it is called by init/calibrate.c when 'ARCH_HAS_READ_CURRENT_TIMER' is
>   defined.
>   since arm64 can implement it, better to provide it as an architect
>   features to let outside use.

Nobody disputes the interest of read_current_timer.

> For the implementation of read_current_timer():
> 
>   it has to face various configurations
>     (e.g. CONFIG_ARM_ARCH_TIMER, arch_timer_read_zero,
>     arch_counter_get_cntvct, arch_counter_get_cntpct)
>   so better still use variable instead of.
>     (excuse me, I do not know what is 'CNTVCT_EL0', is it like a
constant
>     number ?)

Architected timer is mandatory on arm64, so we can always rely on it it be
present. CNTVCT_EL0 is the system register accessing the Virtual Counter,
which is basically what read_current_timer() returns.

> For the implementation of get_cycles()
> 
>   if read_current_timer() is provided,
>   better to let get_cycles() to call it, instead of implement once
again.

There is certainly some value in reusing existing code, but in this
particular case we can simply inline two instructions (isb + mrs
cntvct_el0), and I'm not even completely sure about the isb.

        M.
-- 
Fast, cheap, reliable. Pick two.