[PATCH] ARM: OMAP2+: omap-device: remove omap_device_late_init call completely

[Grygorii Strashko]

Hi Tony,

On 08/26/2015 09:10 PM, Tony Lindgren wrote:
> * Grygorii Strashko <grygorii.strashko at ti.com> [150826 11:01]:
>> Now Kernel fails to boot 50% of times (form build to build) with
>> RT-patchset applied due to the following race - on late boot
>> stages deferred_probe_work_func races with omap_device_late_ini
>>
>> late_initcall
>>   - deferred_probe_initcal() tries to re-probe all pending driver's probe.
>>     [In general, It's NOT expected to probe any other built-in drivers after
>>     deferred_probe_initcal() is finished, because most of
>>     late_initcall_sync/late_initcall functions expected that all driver
>>     or probed or deferred already.]
>>
>> - later on, some driver is probing in this case It's could cpsw.c
>>    (but could be any other drivers)
>>    cpsw_init
>>    - platform_driver_register
>>      - really_probe
>>         - driver_bound
>>           - driver_deferred_probe_trigger
>>    and boot proceed.
>>    So, at this moment we have  deferred_probe_work_func scheduled.
>>
>> late_initcall_sync
>>    - omap_device_late_init
>>      - omap_device_idle
>>
>> CPU1					CPU2
>>    - deferred_probe_work_func
>>      - really_probe
>>        - omap_hsmmc_probe
>> 	- pm_runtime_get_sync
>> 					late_initcall_sync
>> 					- omap_device_late_init
>> 						if (od->_driver_status != BUS_NOTIFY_BOUND_DRIVER) {
>> 							if (od->_state == OMAP_DEVICE_STATE_ENABLED) {
>> 								- omap_device_idle [ops - IP is disabled, ]
>> 	- [fail]
>> 	- pm_runtime_put_sync
>>            - omap_hsmmc_runtime_suspend [ooops!]
> 
> OK idling of unclaimed devices should not happen for deferred probe,
> it should only happen when there's no driver and no probing happening.
>   
>> Lets remove just remove omap_device_late_init completely as suggested
>> by Tero Kristo:
>>
>> "How about remove omap_device_late_init call completely. I don't think
>> it does anything useful at the moment; none of the omap devices get
>> enabled outside runtime_pm, so there should be no need to explicitly
>> disable the devices."
> 
> I think this is still needed from PM point of view as otherwise we
> don't idle any devices that don't have a driver available. Or am I
> missing something?
> 
> To me it seems the bug is relying on the BUS_NOTIFY_BOUND_DRIVER is
> not set in the deferred probe case.
> 


What do you think about below alternative?

diff --git a/arch/arm/mach-omap2/omap_device.c b/arch/arm/mach-omap2/omap_device.c
index 4cb8fd9..72ebc4c 100644
--- a/arch/arm/mach-omap2/omap_device.c
+++ b/arch/arm/mach-omap2/omap_device.c
@@ -901,7 +901,8 @@ static int __init omap_device_late_idle(struct device *dev, void *data)
                if (od->hwmods[i]->flags & HWMOD_INIT_NO_IDLE)
                        return 0;
 
-       if (od->_driver_status != BUS_NOTIFY_BOUND_DRIVER) {
+       if (od->_driver_status != BUS_NOTIFY_BOUND_DRIVER &&
+           od->_driver_status != BUS_NOTIFY_BIND_DRIVER) {
                if (od->_state == OMAP_DEVICE_STATE_ENABLED) {
                        dev_warn(dev, "%s: enabled but no driver.  Idling\n",
                                 __func__);



-- 
regards,
-grygorii