Re: [PATCH v9 4/8] PM / devfreq: Move more initialization before registration

[Leonard Crestez <leonard.crestez@nxp.com>]

On 01.11.2019 10:25, Chanwoo Choi wrote:
> On 19. 10. 31. 오후 10:31, Leonard Crestez wrote:
>> On 31.10.2019 05:10, Chanwoo Choi wrote:
>>> Hi Leonard,
>>>
>>> This patch didn't get the acked-by from devfreq maintainer.
>>> I think that we need to discuss this patch with more time.
>>> Also, it is possible to make it as the separate patch
>>> from this series.
>>>
>>> IMHO, if you make the separate patch for this and
>>> resend the separate patch on later, I think that
>>> we can merge the remained patch related to PM_QOS.
>>
>> The devfreq initialization cleanups are required for dev_pm_qos support,
>> otherwise lockdep warnings are triggered. I can post the cleanups as a
>> separate series but the PM QoS one would depend on the cleanups.
>>
>> Do you prefer multiple smaller series?
> 
> After read the v10, I think v9 is better than v10
> for this issue. >>
>> I try to order my patches with uncontroversial fixes and cleanups first
>> so in theory the earlier parts could be applied separately. It's very
>> rare to see series partially applied though.
>>
>> Earlier objection was that devm should be kept, I think this can be
>> accomplished by splitting device_register into device_initialize and
>> device_add.
>>
>>> On 19. 10. 3. 오전 4:25, Leonard Crestez wrote:
>>>> In general it is a better to initialize an object before making it
>>>> accessible externally (through device_register).
>>>>
>>>> This makes it possible to avoid remove locking the partially initialized
>>>> object and simplifies the code. However devm is not available before
>>>> device_register (only after the device_initialize step) so the two
>>>> allocations need to be managed manually.
>>>>
>>>> Signed-off-by: Leonard Crestez <leonard.crestez@nxp.com>
>>>> Reviewed-by: Matthias Kaehlcke <mka@chromium.org>
>>>> ---
>>>>    drivers/devfreq/devfreq.c | 43 +++++++++++++++++++++++----------------
>>>>    1 file changed, 25 insertions(+), 18 deletions(-)
>>>>
>>>> diff --git a/drivers/devfreq/devfreq.c b/drivers/devfreq/devfreq.c
>>>> index 3e0e936185a3..0b40f40ee7aa 100644
>>>> --- a/drivers/devfreq/devfreq.c
>>>> +++ b/drivers/devfreq/devfreq.c
>>>> @@ -591,10 +591,12 @@ static void devfreq_dev_release(struct device *dev)
>>>>    	mutex_unlock(&devfreq_list_lock);
>>>>    
>>>>    	if (devfreq->profile->exit)
>>>>    		devfreq->profile->exit(devfreq->dev.parent);
>>>>    
>>>> +	kfree(devfreq->time_in_state);
>>>> +	kfree(devfreq->trans_table);
>>>>    	mutex_destroy(&devfreq->lock);
>>>>    	kfree(devfreq);
>>>>    }
>>>>    
>>>>    /**
>>>> @@ -674,44 +676,43 @@ struct devfreq *devfreq_add_device(struct device *dev,
>>>>    	devfreq->max_freq = devfreq->scaling_max_freq;
>>>>    
>>>>    	devfreq->suspend_freq = dev_pm_opp_get_suspend_opp_freq(dev);
>>>>    	atomic_set(&devfreq->suspend_count, 0);
>>>>    
>>>> -	dev_set_name(&devfreq->dev, "devfreq%d",
>>>> -				atomic_inc_return(&devfreq_no));
>>>> -	err = device_register(&devfreq->dev);
>>>> -	if (err) {
>>>> -		mutex_unlock(&devfreq->lock);
>>>> -		put_device(&devfreq->dev);
>>>> -		goto err_out;
>>>> -	}
>>>> -
>>>> -	devfreq->trans_table = devm_kzalloc(&devfreq->dev,
>>>> +	devfreq->trans_table = kzalloc(
>>>>    			array3_size(sizeof(unsigned int),
>>>>    				    devfreq->profile->max_state,
>>>>    				    devfreq->profile->max_state),
>>>>    			GFP_KERNEL);
>>>>    	if (!devfreq->trans_table) {
>>>>    		mutex_unlock(&devfreq->lock);
>>>>    		err = -ENOMEM;
>>>> -		goto err_devfreq;
>>>> +		goto err_dev;
>>>>    	}
>>>>    
>>>> -	devfreq->time_in_state = devm_kcalloc(&devfreq->dev,
>>>> -			devfreq->profile->max_state,
>>>> -			sizeof(unsigned long),
>>>> -			GFP_KERNEL);
>>>> +	devfreq->time_in_state = kcalloc(devfreq->profile->max_state,
>>>> +					 sizeof(unsigned long),
>>>> +					 GFP_KERNEL);
>>>>    	if (!devfreq->time_in_state) {
>>>>    		mutex_unlock(&devfreq->lock);
>>>>    		err = -ENOMEM;
>>>> -		goto err_devfreq;
>>>> +		goto err_dev;
>>>>    	}
>>>>    
>>>>    	devfreq->last_stat_updated = jiffies;
>>>>    
>>>>    	srcu_init_notifier_head(&devfreq->transition_notifier_list);
>>>>    
>>>> +	dev_set_name(&devfreq->dev, "devfreq%d",
>>>> +				atomic_inc_return(&devfreq_no));
>>>> +	err = device_register(&devfreq->dev);
>>>> +	if (err) {
>>>> +		mutex_unlock(&devfreq->lock);
>>>> +		put_device(&devfreq->dev);
>>>> +		goto err_out;
> 
> err_out -> err_dev
> When failed to register, have to free resource.

According documentation on device_register resources need to be freed 
via put_device (which calls dev.release = devfreq_dev_release).

This chunk isn't new; it just appears so because other code was moved 
around it.

> 
>>>> +	}
>>>> +
>>>>    	mutex_unlock(&devfreq->lock);
>>>>    
>>>>    	mutex_lock(&devfreq_list_lock);
>>>>    
>>>>    	governor = try_then_request_governor(devfreq->governor_name);
>>>> @@ -737,14 +738,20 @@ struct devfreq *devfreq_add_device(struct device *dev,
>>>>    
>>>>    	return devfreq;
>>>>    
>>>>    err_init:
>>>>    	mutex_unlock(&devfreq_list_lock);
>>>> -err_devfreq:
>>>>    	devfreq_remove_device(devfreq);
>>>> -	devfreq = NULL;
>>>> +	return ERR_PTR(err);
> 
> 
> It is not proper to return on the middle
> of the exception handling. Need to consider more clean method.

Current code already does "devfreq = NULL" in order to skip the kfree 
below. There are already many cleanup/exit paths, it's just not very 
obvious.

Cleanup for errors before and after "device_register" is different and 
this is required by device core.

This can be improved by splitting device_register into device_initialize 
and device_add: this means that all early error paths do put_device and 
there is no longer any need to manually kfree(devfreq) because 
devfreq_dev_release is called on all cleanup paths. It also preserves devm:

https://patchwork.kernel.org/patch/11221873/

However there is still a difference between cleanup before/after 
device_add: if something fails after device_add you need to call 
devfreq_remove_device (which does device_unregister) and otherwise 
device_put.
>>>> +
>>>>    err_dev:
>>>> +	/*
>>>> +	 * Cleanup path for errors that happen before registration.
>>>> +	 * Otherwise we rely on devfreq_dev_release.
>>>> +	 */
>>>> +	kfree(devfreq->time_in_state);
>>>> +	kfree(devfreq->trans_table);
>>>>    	kfree(devfreq);
>>>>    err_out:
>>>>    	return ERR_PTR(err);
>>>>    }
>>>>    EXPORT_SYMBOL(devfreq_add_device);