Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1660276 > unrolled thread

Re: [PATCH 1/3] nvmem: core: remove member users from struct nvmem_device

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [PATCH 1/3] nvmem: core: remove member users from struct  nvmem_device Heiner Kallweit <hkallweit1@gmail.com> - 2017-06-08 00:00 +0200
    Re: [PATCH 1/3] nvmem: core: remove member users from struct  nvmem_device Srinivas Kandagatla <srinivas.kandagatla@linaro.org> - 2017-06-08 08:40 +0200

#1660276 — Re: [PATCH 1/3] nvmem: core: remove member users from struct nvmem_device

Am 07.06.2017 um 17:30 schrieb Srinivas Kandagatla:
> 
> 
> On 04/06/17 12:01, Heiner Kallweit wrote:
>> Member users is used only to check whether we're allowed to remove
>> the module. So in case of built-in it's not used at all and in case
> 
> nvmem providers doesn't have to be independent drivers, providers could be part of the other driver which can dynamically register and unregister nvmem providers. For example at24 and at25 drivers.
> 
> This patch will break such cases !!
> 
Thanks for the quick review.
I don't think this patch breaks e.g. at24 / at25. Let me try to explain:

at24 / at25 set themself as owner in struct nvmem_device and nvmem_unregister
is called from at24/25_remove only. These remove callbacks are called only if
all references to the respective module have been released.

In current kernel code I don't see any nvmem use broken by the proposed patch.
However in general you're right, there may be future use cases where
nvmem_unregister isn't called only from a remove callback.

If the refcount isn't zero when calling nvmem_unregister then there's a bigger
problem, I don't think there's any normal use case where this can happen.
Instead of just returning -EBUSY I think a WARN() would be appropriate.
Currently no caller of nvmem_unregister checks the return code anyway.
My opinion would be that the refcount here is more a debug feature.


Whilst we're talking about nvmem_unregister:
I think the device_del() at the end should be a device_unregister().
Else we miss put_device as second part of destroying a device.

Rgds, Heiner


> 
> 
>> that owner is a module we have the module refcount for the same
>> purpose already. Whenever users is incremented the owner's refcount
>> is incremented too. Therefore users isn't needed.
>>
>> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
>> ---
>>  drivers/nvmem/core.c | 16 ----------------
>>  1 file changed, 16 deletions(-)
>>
>> diff --git a/drivers/nvmem/core.c b/drivers/nvmem/core.c
>> index 8c830a80..4e07f3f8 100644
>> --- a/drivers/nvmem/core.c
>> +++ b/drivers/nvmem/core.c
>> @@ -33,7 +33,6 @@ struct nvmem_device {
>>      int            word_size;
>>      int            ncells;
>>      int            id;
>> -    int            users;
>>      size_t            size;
>>      bool            read_only;
>>      int            flags;
>> @@ -517,13 +516,6 @@ EXPORT_SYMBOL_GPL(nvmem_register);
>>   */
>>  int nvmem_unregister(struct nvmem_device *nvmem)
>>  {
>> -    mutex_lock(&nvmem_mutex);
>> -    if (nvmem->users) {
>> -        mutex_unlock(&nvmem_mutex);
>> -        return -EBUSY;
>> -    }
>> -    mutex_unlock(&nvmem_mutex);
>> -
>>      if (nvmem->flags & FLAG_COMPAT)
>>          device_remove_bin_file(nvmem->base_dev, &nvmem->eeprom);
>>
>> @@ -562,7 +554,6 @@ static struct nvmem_device *__nvmem_device_get(struct device_node *np,
>>          }
>>      }
>>
>> -    nvmem->users++;
>>      mutex_unlock(&nvmem_mutex);
>>
>>      if (!try_module_get(nvmem->owner)) {
>> @@ -570,10 +561,6 @@ static struct nvmem_device *__nvmem_device_get(struct device_node *np,
>>              "could not increase module refcount for cell %s\n",
>>              nvmem->name);
>>
>> -        mutex_lock(&nvmem_mutex);
>> -        nvmem->users--;
>> -        mutex_unlock(&nvmem_mutex);
>> -
>>          return ERR_PTR(-EINVAL);
>>      }
>>
>> @@ -583,9 +570,6 @@ static struct nvmem_device *__nvmem_device_get(struct device_node *np,
>>  static void __nvmem_device_put(struct nvmem_device *nvmem)
>>  {
>>      module_put(nvmem->owner);
>> -    mutex_lock(&nvmem_mutex);
>> -    nvmem->users--;
>> -    mutex_unlock(&nvmem_mutex);
>>  }
>>
>>  static int nvmem_match(struct device *dev, void *data)
>>
> 

[toc] | [next] | [standalone]

#1660787

On 07/06/17 22:51, Heiner Kallweit wrote:
> Am 07.06.2017 um 17:30 schrieb Srinivas Kandagatla:
>>
>>
>> On 04/06/17 12:01, Heiner Kallweit wrote:
>>> Member users is used only to check whether we're allowed to remove
>>> the module. So in case of built-in it's not used at all and in case
>>
>> nvmem providers doesn't have to be independent drivers, providers could be part of the other driver which can dynamically register and unregister nvmem providers. For example at24 and at25 drivers.
>>
>> This patch will break such cases !!
>>
> Thanks for the quick review.
> I don't think this patch breaks e.g. at24 / at25. Let me try to explain:
> 
> at24 / at25 set themself as owner in struct nvmem_device and nvmem_unregister
> is called from at24/25_remove only. These remove callbacks are called only if
> all references to the respective module have been released.
> 
> In current kernel code I don't see any nvmem use broken by the proposed patch.
> However in general you're right, there may be future use cases where
> nvmem_unregister isn't called only from a remove callback.

Yes, the patch would not break the exiting code, but as said it would 
break a feature which was considered while writing the code.

> 
> If the refcount isn't zero when calling nvmem_unregister then there's a bigger
> problem, I don't think there's any normal use case where this can happen.

Yes I understand chances of this error path is slim but it would crash 
the system if it hits this path, so this safety check is in place.

> Instead of just returning -EBUSY I think a WARN() would be appropriate.
> Currently no caller of nvmem_unregister checks the return code anyway.
> My opinion would be that the refcount here is more a debug feature.
> 
> 
> Whilst we're talking about nvmem_unregister:
> I think the device_del() at the end should be a device_unregister().
> Else we miss put_device as second part of destroying a device.

These issues have already been addressed in
https://patchwork.kernel.org/patch/9685559/
https://patchwork.kernel.org/patch/9685561/
https://patchwork.kernel.org/patch/9729235/


--srini

> 
> Rgds, Heiner
> 
> 
>>
>>
>>> that owner is a module we have the module refcount for the same
>>> purpose already. Whenever users is incremented the owner's refcount
>>> is incremented too. Therefore users isn't needed.
>>>
>>> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
>>> ---
>>>   drivers/nvmem/core.c | 16 ----------------
>>>   1 file changed, 16 deletions(-)
>>>
>>> diff --git a/drivers/nvmem/core.c b/drivers/nvmem/core.c
>>> index 8c830a80..4e07f3f8 100644
>>> --- a/drivers/nvmem/core.c
>>> +++ b/drivers/nvmem/core.c
>>> @@ -33,7 +33,6 @@ struct nvmem_device {
>>>       int            word_size;
>>>       int            ncells;
>>>       int            id;
>>> -    int            users;
>>>       size_t            size;
>>>       bool            read_only;
>>>       int            flags;
>>> @@ -517,13 +516,6 @@ EXPORT_SYMBOL_GPL(nvmem_register);
>>>    */
>>>   int nvmem_unregister(struct nvmem_device *nvmem)
>>>   {
>>> -    mutex_lock(&nvmem_mutex);
>>> -    if (nvmem->users) {
>>> -        mutex_unlock(&nvmem_mutex);
>>> -        return -EBUSY;
>>> -    }
>>> -    mutex_unlock(&nvmem_mutex);
>>> -
>>>       if (nvmem->flags & FLAG_COMPAT)
>>>           device_remove_bin_file(nvmem->base_dev, &nvmem->eeprom);
>>>
>>> @@ -562,7 +554,6 @@ static struct nvmem_device *__nvmem_device_get(struct device_node *np,
>>>           }
>>>       }
>>>
>>> -    nvmem->users++;
>>>       mutex_unlock(&nvmem_mutex);
>>>
>>>       if (!try_module_get(nvmem->owner)) {
>>> @@ -570,10 +561,6 @@ static struct nvmem_device *__nvmem_device_get(struct device_node *np,
>>>               "could not increase module refcount for cell %s\n",
>>>               nvmem->name);
>>>
>>> -        mutex_lock(&nvmem_mutex);
>>> -        nvmem->users--;
>>> -        mutex_unlock(&nvmem_mutex);
>>> -
>>>           return ERR_PTR(-EINVAL);
>>>       }
>>>
>>> @@ -583,9 +570,6 @@ static struct nvmem_device *__nvmem_device_get(struct device_node *np,
>>>   static void __nvmem_device_put(struct nvmem_device *nvmem)
>>>   {
>>>       module_put(nvmem->owner);
>>> -    mutex_lock(&nvmem_mutex);
>>> -    nvmem->users--;
>>> -    mutex_unlock(&nvmem_mutex);
>>>   }
>>>
>>>   static int nvmem_match(struct device *dev, void *data)
>>>
>>
> 

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web