[PATCH v2 2/3] spi: pxa2xx: Prepare for edge-triggered interrupts

Robert Jarzmik robert.jarzmik at free.fr
Wed Jan 18 00:21:35 PST 2017


Jan Kiszka <jan.kiszka at siemens.com> writes:

> On 2017-01-17 08:54, Robert Jarzmik wrote:
>> Jan Kiszka <jan.kiszka at siemens.com> writes:
>> 
>>> When using the a device with edge-triggered interrupts, such as MSIs,
>>> the interrupt handler has to ensure that there is a point in time during
>>> its execution where all interrupts sources are silent so that a new
>>> event can trigger a new interrupt again.
>>>
>>> This is achieved here by looping over SSSR evaluation. We need to take
>>> into account that SSCR1 may be changed by the transfer handler, thus we
>>> need to redo the mask calculation, at least regarding the volatile
>>> interrupt enable bit (TIE).
>>>
>>> Signed-off-by: Jan Kiszka <jan.kiszka at siemens.com>
>> Hi Jan,
>> 
>>> +	while (1) {
>> This bit worries me a bit, as this can be either :
>>  - hogging the SoC's CPU, endlessly running
>>  - or even worse, blocking the CPU for ever
>> 
>> The question behind is, should this be done in a top-half, or moved to a irq
>> thread ?
>
> Every device with a broken interrupt source can hog CPUs, nothing
> special with this one. If you don't close the loop in the handler
> itself, you close it over the hardware retriggering the interrupt over
> and over again.
I'm not speaking of a broken interrupt source, I'm speaking of a broken code,
such as in the handler, or broken status readback, or lack of understanding on
the status register which may imply the while(1) to loop forever.

> So, I don't see a point in offloading to a thread. The normal case is
> some TX done (FIFO available) event followed by an RX event, then the
> transfer is complete, isn't it?
The point is if you stay forever in the while(1) loop, you can at least have a
print a backtrace (LOCKUP_DETECTOR).

>> Imagine that on first iteration the handler returns IRQ_NONE, and on second
>> IRQ_HANDLED. This makes ret IRQ_HANDLED. Yet after the first iteration the
>> handler should have exited, especially if the interrupt is shared with another
>> driver.
>
> That would be a bug in transfer_handler, because we don't enter it
> without a reason (status != 0).
Sure, but can you be sure that all the people modifying the code after you will
see that also ? The other way will _force_ them to see it.

>> Another thing which is along what Andy already said : it would be better
>> practice to have this loop in the form :
>> do {
>> ...
>> } while (exit_condition_not_met);
>
> This implies code duplication in order to calculate the condition
> (mask...). I can do this if desired, I wouldn't do this to my own code,
> though.
Okay, that's acceptable.
Why not have something like this :

sccr1_reg = pxa2xx_spi_read(drv_data, SSCR1);
if (!(sccr1_reg & SSCR1_TIE))
	mask &= ~SSSR_TFS;

/* Ignore RX timeout interrupt if it is disabled */
if (!(sccr1_reg & SSCR1_TINTE))
	mask &= ~SSSR_TINT;

status = pxa2xx_spi_read(drv_data, SSR);
while (status & mask) {
   ... handlers etc ...
   status = pxa2xx_spi_read(drv_data, SSR);
};

There is a duplication of the status read, but that looks acceptable, and the
mask calculation is moved out of the loop (this should be checked more
thoroughly as it looked to me only probe() would change these values, yet I
might be wrong).

Cheers.

-- 
Robert



More information about the linux-arm-kernel mailing list