Re: [PATCH 1/3] net: ethernet: ti: cpdma: fix lockup in cpdma_ctlr_destroy()

[Grygorii Strashko]

On 07/26/2016 11:54 PM, ivan.khoronzhuk wrote:
> 
> 
> On 26.07.16 19:02, Grygorii Strashko wrote:
>> On 07/23/2016 09:24 AM, Ivan Khoronzhuk wrote:
>>>
>>>
>>> On 22.07.16 16:58, Grygorii Strashko wrote:
>>>> Fix deadlock in cpdma_ctlr_destroy() which is triggered now on
>>>> cpsw module removal:
>>>>  cpsw_remove()
>>>>  - cpdma_ctlr_destroy()
>>>>    - spin_lock_irqsave(&ctlr->lock, flags)
>>>>    - cpdma_ctlr_stop()
>>>>      - spin_lock_irqsave(&ctlr->lock, flags); <- deadlock
>>>>    - cpdma_chan_destroy()
>>>>      - spin_lock_irqsave(&ctlr->lock, flags); <- deadlock
>>>>
>>>> The issue has not been observed before because CPDMA channels have
>>>> been destroyed manually by CPSW until commit d941ebe88a41 ("net:
>>>> ethernet: ti: cpsw: use destroy ctlr to destroy channels") was merged.
>>>>
>>>> Signed-off-by: Grygorii Strashko <grygorii.strashko@xxxxxx>
>>>> ---
>>>>  drivers/net/ethernet/ti/davinci_cpdma.c | 2 --
>>>>  1 file changed, 2 deletions(-)
>>>>
>>>> diff --git a/drivers/net/ethernet/ti/davinci_cpdma.c
>>>> b/drivers/net/ethernet/ti/davinci_cpdma.c
>>>> index a68652a..89242e9 100644
>>>> --- a/drivers/net/ethernet/ti/davinci_cpdma.c
>>>> +++ b/drivers/net/ethernet/ti/davinci_cpdma.c
>>>> @@ -436,7 +436,6 @@ int cpdma_ctlr_destroy(struct cpdma_ctlr *ctlr)
>>>>      if (!ctlr)
>>>>          return -EINVAL;
>>>>
>>>> -    spin_lock_irqsave(&ctlr->lock, flags);
>>> Should ctlr->state be checked under lock?
>>> Seems like here should be used unlocked static versions of
>>> cpdma_ctlr_stop() and cpdma_chan_destroy() instead.
>>
>> As per my understanding it's not expected the ctlr->state will be
>> changed at this
>> moment as all net devices has been unregistered already.
> Seems yes, the race can be only in case of incorrect usage, stop while
> destroy,
> destroy while start...etc..all they are mostly unreal use-cases, you are
> right,
> but such check w/o lock always under eyes control, that always makes you
> think
> that smth wrong.
> 
>>
>>>
>>>>      if (ctlr->state != CPDMA_STATE_IDLE)
>>
>> May be I can move above check in cpdma_ctlr_stop() instead.
>> What do you think?
> Yes, it be more clear.
> I was thinking about lock deletion also, as under this destroy function the
> ctlr destroys it's resources one by one, ok, the channels are destroyed
> under lock,
> but pool ....(it's good that it's destroyed after channels). I see that
> it should never
> happen, but ctrl is external structure, who knows as it can be used
> while destroying.
> That was my paranoiac point, so don't pay a lot attention to it. In case
> of normal usage,
> as it's currently is and should be, the lock can be removed.


I'm going to keep it as is after some thinking and code checking -
I don't see any reasons for races here and I can't simply move this check in cpdma_ctlr_stop() 
as it might break ndo_open failure handling (and this is not smth. I'd like to fix within this series).

I'll resend v2 with build issue fixed and with fix for new issue I've found.
> 
>>
>>>>          cpdma_ctlr_stop(ctlr);
>>>>
>>>> @@ -444,7 +443,6 @@ int cpdma_ctlr_destroy(struct cpdma_ctlr *ctlr)
>>>>          cpdma_chan_destroy(ctlr->channels[i]);
>>>>
>>>>      cpdma_desc_pool_destroy(ctlr->pool);
>>>> -    spin_unlock_irqrestore(&ctlr->lock, flags);
>>>>      return ret;
>>>>  }
>>>>  EXPORT_SYMBOL_GPL(cpdma_ctlr_destroy);
>>>>
>>>
>>
>>


-- 
regards,
-grygorii