Re: [PATCH] i2c: iproc: fix race between client unreg and isr

[Ray Jui <ray.jui@broadcom.com>]

On 7/22/2020 3:41 AM, Wolfram Sang wrote:
> 
>>> +	synchronize_irq(iproc_i2c->irq);
>>
>> If one takes a look at the I2C slave ISR routine, there are places where
>> IRQ can be re-enabled in the ISR itself. What happens after we mask all
>> slave interrupt and when 'synchronize_irq' is called, which I suppose is
>> meant to wait for inflight interrupt to finish where there's a chance
>> the interrupt can be re-enable again? How is one supposed to deal with that?
> 
> I encountered the same problem with the i2c-rcar driver before I left
> for my holidays.
> 

I think the following sequence needs to be implemented to make this
safe, i.e., after 'synchronize_irq', no further slave interrupt will be
fired.

In 'bcm_iproc_i2c_unreg_slave':

1. Set an atomic variable 'unreg_slave' (I'm bad in names so please come
up with a better name than this)

2. Disable all slave interrupts

3. synchronize_irq

4. Set slave to NULL

5. Erase slave addresses

In the ISR routine, it should always check against 'unreg_slave' before
enabling any slave interrupt. If 'unreg_slave' is set, no slave
interrupt should be re-enabled from within the ISR.

I think the above sequence can ensure no further slave interrupt after
'synchronize_irq'. I suggested using an atomic variable instead of
variable + spinlock due to the way how sync irq works, i.e., "If you use
this function while holding a resource the IRQ handler may need you will
deadlock.".

Thanks,

Ray

>>> +	iproc_i2c->slave = NULL;
>>> +
>>>  	/* Erase the slave address programmed */
>>>  	tmp = iproc_i2c_rd_reg(iproc_i2c, S_CFG_SMBUS_ADDR_OFFSET);
>>>  	tmp &= ~BIT(S_CFG_EN_NIC_SMB_ADDR3_SHIFT);
>>>