Re: [PATCH] crypto: picoxcell: add missed tasklet_kill

[Marc Gonzalez]

On 22/11/2019 09:55, Herbert Xu wrote:

> On Fri, Nov 15, 2019 at 10:31:16AM +0800, Chuhong Yuan wrote:
>> This driver forgets to kill tasklet when probe fails and remove.
>> Add the calls to fix it.
>>
>> Signed-off-by: Chuhong Yuan <hslester96@xxxxxxxxx>
> 
> Yes this driver does look buggy but I think your patch isn't
> enough.
> 
>> diff --git a/drivers/crypto/picoxcell_crypto.c b/drivers/crypto/picoxcell_crypto.c
>> index 3cbefb41b099..8d7c6bb2876e 100644
>> --- a/drivers/crypto/picoxcell_crypto.c
>> +++ b/drivers/crypto/picoxcell_crypto.c
>> @@ -1755,6 +1755,7 @@ static int spacc_probe(struct platform_device *pdev)
>>  	if (!ret)
>>  		return 0;
>>  
>> +	tasklet_kill(&engine->complete);
> 
> The tasklet is schedule by the IRQ handler so you should not kill
> it until the IRQ handler has been unregistered.
> 
> This driver is also buggy because it registers the IRQ handler
> before initialising the tasklet.  You must always be prepared for
> spurious IRQs.  IOW, as soon as you register the IRQ handler you
> must be prepared for it to be called.
> 
>> @@ -1771,6 +1772,7 @@ static int spacc_remove(struct platform_device *pdev)
>>  	struct spacc_alg *alg, *next;
>>  	struct spacc_engine *engine = platform_get_drvdata(pdev);
>>  
>> +	tasklet_kill(&engine->complete);
> 
> Ditto.
> 
> However, the IRQ handler is registered through devm which makes it
> hard to kill the tasklet after unregistering it.  We should probably
> convert it to a normal request_irq so we can control how it's
> unregistered.

Or inversely, registering the tasklet_kill() through devm, so that it
is called *after* the ISR unregistration.

Regards.