Re: [PATCH v12 7/9] firmware: qcom: scm: Fix __scm->dev assignement

[Mukesh Ojha <quic_mojha@quicinc.com>]

On 3/19/2024 3:52 PM, Pavan Kondeti wrote:
> On Tue, Mar 19, 2024 at 03:38:57PM +0530, Mukesh Ojha wrote:
>>
>>
>> On 3/19/2024 6:47 AM, Pavan Kondeti wrote:
>>> On Mon, Mar 18, 2024 at 06:38:20PM +0530, Mukesh Ojha wrote:
>>>>
>>>>
>>>> On 3/3/2024 12:55 AM, Bjorn Andersson wrote:
>>>>> On Tue, Feb 27, 2024 at 09:23:06PM +0530, Mukesh Ojha wrote:
>>>>>> qcom_scm_is_available() gives wrong indication if __scm
>>>>>> is initialized but __scm->dev is not.
>>>>>>
>>>>>> Fix this appropriately by making sure if __scm is
>>>>>> initialized and then it is associated with its
>>>>>> device.
>>>>>>
>>>>>
>>>>> This seems like a bug fix, and should as such have a Fixes: tag and
>>>>> probably Cc: stable@vger.kernel.org
>>>>>
>>>>>> Signed-off-by: Mukesh Ojha <quic_mojha@quicinc.com>
>>>>>> ---
>>>>>>     drivers/firmware/qcom/qcom_scm.c | 2 +-
>>>>>>     1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>>
>>>>>> diff --git a/drivers/firmware/qcom/qcom_scm.c b/drivers/firmware/qcom/qcom_scm.c
>>>>>> index 6c252cddd44e..6f14254c0c10 100644
>>>>>> --- a/drivers/firmware/qcom/qcom_scm.c
>>>>>> +++ b/drivers/firmware/qcom/qcom_scm.c
>>>>>> @@ -1859,6 +1859,7 @@ static int qcom_scm_probe(struct platform_device *pdev)
>>>>>>     	if (!scm)
>>>>>>     		return -ENOMEM;
>>>>>> +	scm->dev = &pdev->dev;
>>>>>>     	ret = qcom_scm_find_dload_address(&pdev->dev, &scm->dload_mode_addr);
>>>>>>     	if (ret < 0)
>>>>>>     		return ret;
>>>>>> @@ -1895,7 +1896,6 @@ static int qcom_scm_probe(struct platform_device *pdev)
>>>>>>     		return ret;
>>>>>>     	__scm = scm;
>>>>>> -	__scm->dev = &pdev->dev;
>>>>>
>>>>> Is it sufficient to just move the line up, or do we need a barrier of
>>>>> some sort here?
>>>>
>>>> Would be good to use, smp_mb() before the assignment
>>>>        __scm = scm
>>>> along with moving below line
>>>> __scm->dev = &pdev->dev
>>>>
>>>
>>> Full memory barrier is not needed here. store variant is sufficient.
>>> WRITE_ONCE() + smp_store_release() will fit here no?
>>
>> Thanks for the comment, i again have a look at it and agree we don't
>> need a full barrier here.
>>
>> And we can do either of the below two ways.
>>
>> -Mukesh
>>
>>
>> // 1st way
>>
>> diff --git a/drivers/firmware/qcom/qcom_scm.c
>> b/drivers/firmware/qcom/qcom_scm.c
>> index 49ddbcab0680..b638fb407fc6 100644
>> --- a/drivers/firmware/qcom/qcom_scm.c
>> +++ b/drivers/firmware/qcom/qcom_scm.c
>> @@ -1741,7 +1741,12 @@ static int qcom_scm_qseecom_init(struct qcom_scm
>> *scm)
>>    */
>>   bool qcom_scm_is_available(void)
>>   {
>> -       return !!__scm;
>> +       bool avail;
>>    */
>>   bool qcom_scm_is_available(void)
>>   {
>> -       return !!__scm;
>> +       bool avail;
>> +
>> +       avail = !!READ_ONCE(__scm);
>> +       smp_rmb();
>> +
>> +       return avail;
>>   }
>>   EXPORT_SYMBOL_GPL(qcom_scm_is_available);
>>
> 
> Your original problem statement: qcom_scm_is_available() gives wrong indication
> if __scm is initialized but __scm->dev is not.
> 
> This does not require read side barrier as there is an address
> dependency. If the writer does it *correctly*, the reader would always
> observe __scm->dev != NULL when __scm != NULL without any barrier.

It looks like write barrier pairs with an address-dependency barrier, a
control dependency, an acquire barrier, a release barrier, a read 
barrier, or a general barrier.

So, smp_rmb() is redundant here.

Also, for correction, we may not need smp_load_acquire() in the 1st way
and just using READ_ONCE() is enough.

-Mukesh
> 
> Thanks,
> Pavan