Re: [PATCH v2 2/2] mailbox: add STMicroelectronics STM32 IPCC driver

[Fabien DESSENNE]

On 06/04/18 14:56, Jassi Brar wrote:
> On Fri, Apr 6, 2018 at 5:59 PM, Fabien DESSENNE <fabien.dessenne@xxxxxx> wrote:
>> Hi
>>
>>
>> On 05/04/18 11:38, Jassi Brar wrote:
>>> On Mon, Mar 12, 2018 at 4:28 PM, Fabien Dessenne <fabien.dessenne@xxxxxx> wrote:
>>> ....
>>>> +
>>>> +       /* irq */
>>>> +       for (i = 0; i < IPCC_IRQ_NUM; i++) {
>>>> +               ipcc->irqs[i] = of_irq_get_byname(dev->of_node, irq_name[i]);
>>>> +               if (ipcc->irqs[i] < 0) {
>>>> +                       dev_err(dev, "no IRQ specified %s\n", irq_name[i]);
>>>> +                       ret = ipcc->irqs[i];
>>>> +                       goto err_clk;
>>>> +               }
>>>> +
>>>> +               ret = devm_request_threaded_irq(dev, ipcc->irqs[i], NULL,
>>>> +                                               irq_thread[i], IRQF_ONESHOT,
>>>> +                                               dev_name(dev), ipcc);
>>>>
>>> In your interrupt handlers you don't do anything that could block.
>>> Threads only adds some delay to your message handling.
>>> So maybe use devm_request_irq() ?
>> The interrupt handlers call mbox_chan_received_data() /
>> mbox_chan_txdone(), which call in turn client's rx_callback() /
>> tx_done() / tx_prepare() which behavior may be unsafe. Hence, using a
>> threaded irq here seems to be a good choice.
>>
> rx_callback() is supposed to be atomic.

I am worried with this atomic part (and honestly I did not note that the 
callbacks were expected to be)

In my case, remoteproc->virtio->rpmsg is the mailbox client defining the 
rx_callback.
If I follow your suggestion, I shall make this rx_callback Atomic in 
remoteproc (or in virtio or rpmsg). And this does not seem to be so 
simple (add a worker in the middle of somewhere?). Bjorn, feel free to 
comment this part.

An alternate implementation consists in using a threaded IRQ for the 
mailbox interrupt.
This option is not only simple, but also ensures to split bottom & half 
parts at the irq level which is IMHO a general good practice.

I can see that some mailbox clients implement callbacks that are NOT 
atomic and I suspect this is the reason why some mailbox drivers use 
threaded_irq (rockchip mailbox splits the bottom & half parts).

Would it be acceptable to consider the "atomic client callback" as a 
non-strict rule ?

>   So was tx_done() but some
> platforms needed preparing for the message to be sent. Your client is
> not going to be used by other platforms or even over other
> controllers, so if your prepare is NULL/atomic, you should assume
> tx_done to be atomic and not lose performace. If time comes to fix it,
> we'll move prepare() out of the atomic path.
>
>
>>> .......
>>>> +
>>>> +static struct platform_driver stm32_ipcc_driver = {
>>>> +       .driver = {
>>>> +               .name = "stm32-ipcc",
>>>> +               .owner = THIS_MODULE,
>>>>
>>> No need of owner here these days.
>> OK, I will suppress it.
>>
>>> And also maybe use readl/writel, instead of _relaxed.
>> The IPCC device is exclusively used on ARM. In ARM architecture, the
>> ioremap on devices are strictly ordered and uncached.
>> In that case, using _relaxed avoids an unneeded cache flush, slightly
>> improving performance.
>>
> Its not the portability, but that the impact is negligible in favor of
> _relaxed() version when all you do is just program some registers and
> not heavy duty i/o. But I am ok either way.  You'd gain far more
> performance handling irqs in non-threaded manner :)
>
> Cheers!