[alsa-devel] Issue in alsa when dma complete race with pcm release

[Lars-Peter Clausen]

On 07/15/2015 03:37 AM, Shengjiu Wang wrote:
> On Tue, Jul 07, 2015 at 03:32:02PM +0200, Lars-Peter Clausen wrote:
>> On 07/07/2015 12:13 PM, Shengjiu Wang wrote:
>>> On Mon, Jul 06, 2015 at 02:27:01PM +0200, Lars-Peter Clausen wrote:
>>>> On 07/06/2015 11:01 AM, Shengjiu Wang wrote:
>>>>> On Fri, Jul 03, 2015 at 12:56:53PM +0200, Lars-Peter Clausen wrote:
>>>>>> On 07/03/2015 10:25 AM, Shengjiu Wang wrote:
>>>>>>> Hi alsa-devel
>>>>>>>
>>>>>>>     There maybe a issue in ALSA when dma complete race with snd_pcm_release.
>>>>>>> The pcm release and dma complete are in different thread. There is occasion
>>>>>>> that dmaengine_pcm_dma_complete() is called too late, some memory has been
>>>>>>> freed, the prtd is null. Then there is kernel dump.
>>>>>>>
>>>>>>>     Is there any solution for this issue? Thanks.
>>>>>>
>>>>>> We need to introduce a synchronization primitive that allows a
>>>>>> dmaengine client to synchronize to the execution of the complete
>>>>>> callbacks.
>>>>>>
>>>>>> terminate_all() unfortunately can't do this since terminate_all()
>>>>>> might be called from within one of the complete callbacks and so
>>>>>> would cause a deadlock if we'd wait for all complete callbacks to
>>>>>> finish before terminate_all() returns.
>>>>>>
>>>>>> So what is needed is a new function called dmaengine_sync() that
>>>>>> will wait until all scheduled complete callbacks have finished. A
>>>>>> call to this function needs to be put in snd_dmaengine_pcm_close()
>>>>>> before the prtd is closed.
>>>>>>
>>>>>> - Lars
>>>>>
>>>>> How to check " all scheduled complete callbacks have finished"?
>>>>
>>>> That will be up to the dmaengine driver. But it basically comes down
>>>> to two things:
>>>>
>>>> 1) The driver needs to make sure that tasklet_schedule() is no
>>>> longer called after terminate_all() has finished.
>>> Some driver can't make sure this. The dma interrupt may come later after
>>> terminate_all.
>>
>> Most drivers can't make sure that the interrupt routine is not
>> executed after terminate_all() has been called, since both are fully
>> asynchronous. But what the driver needs to take care of is to
>> synchronize the two e.g. using a spin_lock() and make sure that if
>> terminate_all() has been called tasklet_schedule() is not called,
>> even if the isr is executed. Most drivers get this rigght.
>>
>>>> 2) In the sync() callback call tasklet_kill() to make sure that it
>>>> has finished running
>>>>
>>>>>
>>>>> One concern is if add wait in snd_dmaengine_pcm_close(), which wil cause
>>>>> the snd_pcm_release is bound with dmaengine, when there is error in dma
>>>>> and no callback be called. Then the snd_pcm_release will not be released.
>>>>
>>>> The sync() function will only wait if there is a callback scheduled,
>>>> if there is non scheduled it will return immediately.
>>>
>>>
>>> Do you have other method to resolve this issue? or simple method?
>>
>> No, this is the way to go. You have two asynchronous processes and
>> you want to get deterministic execution ordering between the two you
>> need to add a synchronization primitive.
>>
>> - Lars
>>
> Thanks your advice. It is workable. But need to add new api in dmaengine.h
>
> I think about another method, can we add spin_lock in
> dmaengine_pcm_dma_complete() and snd_dmaengine_pcm_close()? and add a flag
> for pcm_close, if pcm closed, then just do nothing in dma complete.
> How do you think for this?

Since you are freeing the memory that would store the flag that wont work.

- Lars