Re: [PATCH] iio:trigger: Fix use_count race condition

[Jonathan Cameron <jic23@kernel.org>]

On 06/11/2013 09:30 PM, Lars-Peter Clausen wrote:
> On 06/11/2013 10:07 PM, Jonathan Cameron wrote:
>> On 06/11/2013 07:18 PM, Lars-Peter Clausen wrote:
>>> When using more than one trigger consumer it can happen that multiple threads
>>> perform a read-modify-update cycle on 'use_count' concurrently. This can cause
>>> updates to be lost and use_count can get stuck at non-zero value, in which case
>>> the IIO core assumes that at least one thread is still running and will wait for
>>> it to finish before running any trigger handlers again. This effectively renders
>>> the trigger disabled and a reboot is necessary before it can be used again. To
>>> fix this make use_count an atomic variable. Also set it to the number of
>>> consumers before starting the first consumer, otherwise it might happen that
>>> use_count drops to 0 even though not all consumers have been run yet.
>>>
>> I am a little worried there is a different race in here. Can't immediateliy get
>> my head around whether it can actually occur. It would require a subirq thread
>> to finish handling the interrupt during either trigger_poll or trigger_poll_chained.
>>
>> I can't immediately see what prevents this happening..
>>
>> One nasty option might be to ensure that we only launch num_consumers interrupts
>> on without caring whether they are the ones we originally counted or not.
>>
>>
>>> Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
>>> ---
>>>  drivers/iio/industrialio-trigger.c | 44 +++++++++++++++++++++++++++-----------
>>>  include/linux/iio/trigger.h        |  3 ++-
>>>  2 files changed, 33 insertions(+), 14 deletions(-)
>>>
>>> diff --git a/drivers/iio/industrialio-trigger.c b/drivers/iio/industrialio-trigger.c
>>> index 4d6c7d8..a02ca65 100644
>>> --- a/drivers/iio/industrialio-trigger.c
>>> +++ b/drivers/iio/industrialio-trigger.c
>>> @@ -126,13 +126,22 @@ static struct iio_trigger *iio_trigger_find_by_name(const char *name,
>>>  
>>>  void iio_trigger_poll(struct iio_trigger *trig, s64 time)
>>>  {
>>> +	unsigned int num_consumers;
>>>  	int i;
>>> -	if (!trig->use_count)
>>> -		for (i = 0; i < CONFIG_IIO_CONSUMERS_PER_TRIGGER; i++)
>>> -			if (trig->subirqs[i].enabled) {
>>> -				trig->use_count++;
>>> +
>>> +	if (!atomic_read(&trig->use_count)) {
>>> +		num_consumers = 0;
>>> +		for (i = 0; i < CONFIG_IIO_CONSUMERS_PER_TRIGGER; i++) {
>>> +			if (trig->subirqs[i].enabled)
>>> +				num_consumers++;
>>> +		}
>>> +		atomic_set(&trig->use_count, num_consumers);
>>> +
>> Is there any chance the state of subirqs[i].enabled might have changed since
>> it was queried above?
> 
> hm, right.
> 
>>
>>> +		for (i = 0; i < CONFIG_IIO_CONSUMERS_PER_TRIGGER; i++) {
>>> +			if (trig->subirqs[i].enabled)
>> how about,
>>
>> if (trig->subirqs[i].enabled && num_consumers--)
>> as that would prevent the case of launching too many irq handlers.
> 
> That wouldn't fix the case where the subirq was enabled. use_count would end
> up with a positive value.
Can't say I follow that given I can't see a way we'd end up with fewer enabled
sub irqs on the second pass than the first.  Thus it would be fine as we would
fire use_count subirqs, each of which would then decrement use_count.

> We can make a copy of enabled and use it for both
> loops. Or use a spinlock protecting the triggers subirqs.

I suspect a spin lock is going to be the cleanest solution.

> 
>>
>>>  				generic_handle_irq(trig->subirq_base + i);
>>> -			}
>>> +		}
>>> +	}
>>>  }
>>>  EXPORT_SYMBOL(iio_trigger_poll);
>>>  
>>> @@ -145,20 +154,29 @@ EXPORT_SYMBOL(iio_trigger_generic_data_rdy_poll);
>>>  
>>>  void iio_trigger_poll_chained(struct iio_trigger *trig, s64 time)
>>>  {
>>> +	unsigned int num_consumers;
>>>  	int i;
>>> -	if (!trig->use_count)
>>> -		for (i = 0; i < CONFIG_IIO_CONSUMERS_PER_TRIGGER; i++)
>>> -			if (trig->subirqs[i].enabled) {
>>> -				trig->use_count++;
>>> -				handle_nested_irq(trig->subirq_base + i);
>>> -			}
>>> +
>>> +	if (!atomic_read(&trig->use_count)) {
>>> +		num_consumers = 0;
>>> +		for (i = 0; i < CONFIG_IIO_CONSUMERS_PER_TRIGGER; i++) {
>>> +			if (trig->subirqs[i].enabled)
>>> +				num_consumers++;
>>> +		}
>>> +		atomic_set(&trig->use_count, num_consumers);
>>> +
>>> +		for (i = 0; i < CONFIG_IIO_CONSUMERS_PER_TRIGGER; i++) {
>>> +			if (trig->subirqs[i].enabled)
>>> +				generic_handle_irq(trig->subirq_base + i);
>>> +		}
>>> +	}
>>>  }
>>>  EXPORT_SYMBOL(iio_trigger_poll_chained);
>>>  
>>>  void iio_trigger_notify_done(struct iio_trigger *trig)
>>>  {
>>> -	trig->use_count--;
>>> -	if (trig->use_count == 0 && trig->ops && trig->ops->try_reenable)
>>> +	if (atomic_dec_and_test(&trig->use_count) && trig->ops &&
>>> +		trig->ops->try_reenable)
>>>  		if (trig->ops->try_reenable(trig))
>>>  			/* Missed an interrupt so launch new poll now */
>>>  			iio_trigger_poll(trig, 0);
>>> diff --git a/include/linux/iio/trigger.h b/include/linux/iio/trigger.h
>>> index 3869c52..369cf2c 100644
>>> --- a/include/linux/iio/trigger.h
>>> +++ b/include/linux/iio/trigger.h
>>> @@ -8,6 +8,7 @@
>>>   */
>>>  #include <linux/irq.h>
>>>  #include <linux/module.h>
>>> +#include <linux/atomic.h>
>>>  
>>>  #ifndef _IIO_TRIGGER_H_
>>>  #define _IIO_TRIGGER_H_
>>> @@ -61,7 +62,7 @@ struct iio_trigger {
>>>  
>>>  	struct list_head		list;
>>>  	struct list_head		alloc_list;
>>> -	int use_count;
>>> +	atomic_t			use_count;
>>>  
>>>  	struct irq_chip			subirq_chip;
>>>  	int				subirq_base;
>>>
>